Sankar Kumar
Palaniappan

Problem Statement

Most "AI data copilot" demos ask an LLM to do everything, including things LLMs are bad at, like guaranteeing a SQL query is safe or reliably counting nulls. Enterprises don't just need a plausible-sounding answer about their data; they need to trust it, and in regulated industries, prove that trust to an auditor.

Solution

An agentic platform built around one rule: deterministic engines own the ground truth, the LLM only explains, classifies, and narrates. Four agents each pair a real profiling, validation, or detection engine with an LLM interpretation layer on top: a Data Quality Agent (ydata-profiling / Great Expectations) for nulls, duplicates, and schema drift; a Catalog Agent (pgvector + metadata graph) for lineage and semantic search; an Analytics Agent (sqlglot) that only ever executes safe, read-only, LIMIT-bound SQL generated from natural language; and a Governance Agent (Microsoft Presidio) that scans for PII/PHI and maps findings to HIPAA, GDPR, and 21 CFR Part 11. Langflow handles the reasoning for each agent, n8n orchestrates scheduling, routing, and the audit trail, and MCP is the bridge connecting them, with no brittle glue-code HTTP calls. The whole pipeline runs against a synthetic pharma/clinical-trial dataset deliberately seeded with real data-quality issues and PII/PHI fields, so it's inspectable end to end rather than a polished demo.

Problem Statement

Matching a patient to a clinical trial means reading a chart against dozens of free-text inclusion and exclusion criteria: age, diagnosis and stage, biomarkers, recent labs, performance status, prior therapy. The data needed to check them is scattered across structured EHR fields, pathology reports, and clinical notes, so coordinators screen by hand, one patient and one trial at a time, and eligible patients are routinely missed, often because the chart is simply incomplete. Dropping an AI agent into that workflow creates a different risk: in healthcare, an answer that sounds right but isn't backed by the record is worse than no answer at all.

Solution

Agentic Clinical Trial Matching is a live, interactive prototype built against a full AI PRD, designed around one decision: what the model is not allowed to do. A deterministic rules engine decides eligibility, scoring every criterion Met, Not met, Unknown, or Needs review, and every Met must cite the exact source text, so a value that isn't in the document is rejected. The AI agent only gathers evidence and drafts: it runs read-only tools under an 8-call cap and a per-user quota, with prompt-injection scanning and patient scoping, and anything it drafts (referrals, outside-records requests) waits for a clinician. Nothing leaves the system until a clinician approves it.

The PRD became 13 features, including a path-to-eligibility planner with a what-if sandbox, an evidence expiry clock with time travel, a one-page visit brief and plain-language patient handout, a tamper-evident audit log chained with SHA-256 plus replay under today's rules, rule impact previews, a guardrail lab where you switch a control off and watch the attack it was stopping succeed, and an eval harness with ship thresholds and a release gate. Rollout runs shadow mode, then human-in-the-loop, then steady state, with promotion gated by exit criteria computed from evidence instead of a button.

To test it, about 4,200 automated checks run across 19 groups using independent oracles and the real API route handlers; 36 bugs were then put back one at a time to confirm the suite catches each one. A second AI reviewing the build cold found 17 problems: the worst, a "deterministic" check that would have accepted a PD-L1 score of 22 because 22C3 is the name of the assay, not the result. All 17 were fixed, each with a regression check. Everything runs on synthetic data; it is not a clinical tool.

Problem Statement

Clinical trials fail for reasons that are rarely stated plainly. Live ClinicalTrials.gov data puts 34,023 trials at TERMINATED status (5.7% of all 595,847 registered studies), and combined with WITHDRAWN and SUSPENDED, 52,362 trials (8.8%) never reached completion as planned. A random sample of 200 TERMINATED trials shows 8.0% with no stop reason recorded at all and another 3.5% citing only a generic "business decision" with no underlying cause, with 11.5% giving literally zero signal. The remaining 88.5% do name a cause, but most are single-line labels like "low enrollment" or "Covid-19" that don't explain the deeper why without cross-referencing enrollment data, results, and publication history by hand. No existing tool does causal reasoning over this public data: ClinicalTrials.gov and the AACT database expose raw fields with no reasoning, PubMed requires manual cross-referencing, and enterprise tools like Citeline/Trialtrove are comprehensive but priced for institutions and still largely descriptive rather than causal.

Solution

ClinicalAI-TrialAnalysis is a single text input: paste an NCT ID and the backend fetches the ClinicalTrials.gov v2 record, resolves linked PubMed publications via NCBI E-utilities, and runs a structured reasoning pass against an explicit 7-category failure taxonomy: recruitment, efficacy, safety/toxicity, funding/business, operational/protocol, strategic/competitive, futility, with few-shot examples and a hidden chain-of-thought scratchpad reasoning through the evidence before committing to a final ranked answer. Every claim across the four output blocks (Bottom Line, Ranked Hypotheses, Evidence, Guardrail) is tagged Fact, Inference, or Hypothesis, enforced by schema validation rather than prompt instruction alone, so an unlabeled claim can never render. The UX is deliberately frictionless: no accounts, no login, no saved history; the product is fully stateless. A malformed NCT ID is rejected client-side before any network call is even made, a well-formed but nonexistent ID gets an explicit "trial not found" state, and if a trial is COMPLETED or still RECRUITING, the tool states plainly that there's no failure to explain instead of fabricating one. If PubMed enrichment fails, it still returns a trial-data-only analysis with a note rather than failing the whole request.

Problem Statement

Comparing two companies at an investor-grade level (financials, valuation multiples, competitive moats, risk factors, forward outlook) normally means either paying for equity research or spending hours cross-referencing filings, news, and analyst notes yourself. Ask a general chatbot instead and you get a plausible-sounding paragraph with no live data, no citations, and no way to check if any of it is current.

Solution

Company Comparison takes two company names and runs a live, GPT-4o-mini-powered deep research pass with real-time web search, then returns a structured comparison across financials, valuation, moats, SWOT, risks, and outlook, with citations attached to the claims that need them. It's a bring-your-own-key tool by design: the OpenAI key is entered client-side, used only to call OpenAI directly from the browser, and never stored or sent anywhere else, so there's no signup, no backend holding credentials, and no cost to run beyond what the user's own key spends. A single report can take up to a minute since it's doing live research rather than replaying cached training data.

Problem Statement

Business professionals (founders, ops managers, procurement leads) routinely sign NDAs and MSAs without fully understanding what they're agreeing to. Without in-house legal teams, reviewing a single contract takes 90–120 minutes and typically costs $1,500–$3,000 at standard $250–$500/hr legal billing rates, and 43% of SMBs report a commercial dispute that traced back to a misunderstood contract term. Existing options are a poor fit at both ends: enterprise contract-lifecycle tools like DocuSign CLM or Ironclad are built for $50k–$500k contracts, and general tools like ChatGPT produce a plausible-sounding summary with no structured extraction, no page attribution, and no way to tell which parts to trust. Rule-based parsers don't hold up either. A standard confidentiality clause alone shows up in 50+ structurally different phrasings across real contracts, which pushes regex-based extraction past a 30% miss rate.

Solution

ContractIQ extracts the 10 terms that matter most in an NDA and the 12 that matter most in an MSA: parties, confidentiality obligations, liability caps, termination clauses, and so on, using GPT-4o in JSON mode at temperature 0.1, with a few-shot prompt built from hand-labeled NDA and MSA examples. Every term comes back with three things a plain summary won't give you: a page number, a 0–100% confidence score that's color-coded and never hidden even below 50% (just flagged with a non-dismissible warning) and the verbatim source sentence it was pulled from, with an expandable "Why?" that answers the trust question directly instead of asking for blind faith. A grounded chat layer answers plain-English follow-up questions using the full contract text as context on every turn, with a system prompt that treats "I cannot find this in the document" as a correct answer rather than a failure, and a mandatory page citation on every response. Hallucination is treated as the top trust risk end to end: deterministic decoding settings, a JSON-parse retry loop, an automated regression test confirming the model still says "not found" on off-document questions, and a monthly calibration check against a legal-SME-labeled set plus the CUAD dataset (13,000+ annotations across 510 contracts), with a UI warning if miscalibration crosses 15%. The whole build is scoped against explicit production-readiness bars before calling anything launched: ≥88% F1 on NDA extraction, ≥85% on MSA, sub-30-second P95 latency, and a correction rate under 12% in any rolling 7-day window. It was developed end-to-end using Claude Code as the primary development partner, as a direct test of how far an agentic coding workflow can carry a real product, from a 28-page PRD with its own HHH (Helpful/Honest/Harmless) readiness framework all the way to a live, shipped app.

Problem Statement

Leaders asking "how do we move AI from pilot to scale?" usually have scattered pilots, no baseline and no agreed criteria for what to fund next. Generic maturity surveys produce a score but no decision, and prioritisation spreadsheets rarely connect to the readiness gaps that will make a pilot fail.

Solution

Each tool feeds the next. Diagnostic gaps raise warnings on risky use cases in the prioritiser, and the shortlist becomes the content of the workshop agenda, down to a pre-mortem on the top pilot. Scoring is transparent: every formula is shown, feasibility weights are adjustable, and the maturity stage can never run more than one step ahead of the weakest area. An optional life-sciences mode adds GxP validation, 21 CFR Part 11 data integrity and human-oversight questions with matching roadmap actions. Optional Claude features tailor the roadmap and suggest use cases and scores; the organisation name is never sent, each visitor consents once, and every response is validated before it's shown. Built with Next.js 16, React 19, TypeScript and Tailwind. There is no login and no database: work stays in the browser and can be saved to a file.

Problem Statement

Prior authorization review is slow because the evidence is buried in faxes, H&Ps, cardiology clearances, labs and orders. Nurses have to find each fact, check it against layered coverage criteria, and defend the decision later. AI that summarizes without sources is unusable here: every assertion has to be traceable, and the cases that need judgment have to reach a human.

Solution

A guided flow of extraction, criteria matching, review and determination, with a shared case header and audit trail throughout. Each criterion shows its status, the governing policy tier (NCD, then LCD, then payer policy, then MCG) and a clickable citation that opens the actual packet PDF with the cited sentence highlighted. An "Ask about this document" panel answers questions about the packet with links back into the source. A review-tier banner explains in plain language why a case needs a clarification request, an escalation to a medical director, or is ready to approve. The nurse, or for higher-risk cases a medical director, always makes the final call. Built with React 19, Vite and Tailwind v4, with pdf.js for in-browser rendering and highlighting. It runs on synthetic data with scripted behavior and no backend; it is not a clinical tool.

Problem Statement

Commercial property managers and multi-site occupiers currently have two bad options: spreadsheets, which are fast to start but error-prone and blind to upcoming deadlines until it's too late, or enterprise/mid-market lease platforms (Yardi, MRI, Visual Lease, LeaseQuery), which are accounting-first, built for ASC 842/IFRS 16 journal entries, expensive ($25K–$500K+/yr), and slow to implement (weeks to 18 months). Nobody has built a lightweight, compliance-and-deadline-first tool for the "moveable middle": mid-market property managers and franchise/multi-unit occupiers with 20–500 leases who care more about renewal windows, permits, insurance certificates, and jurisdictional obligations than right-of-use asset calculations.

Solution

Rather than building a standalone SaaS product from scratch, this ships the core value first as a Claude Skill: a SKILL.md that turns Claude into a lease compliance assistant, runnable immediately inside claude.ai, Claude Code, Claude Cowork, or the API, with no deployment or implementation project. Given lease documents and a portfolio spreadsheet, the skill abstracts key lease terms into a structured format, tracks critical dates and obligations (renewals, notice-to-vacate windows, insurance certificate expirations, permit renewals), surfaces jurisdiction-specific regulatory risk that pure accounting tools miss (zoning, ESG/emissions disclosure laws like NYC Local Law 97, ADA/fire-code items), generates compliance-ready reports, and answers natural-language portfolio questions, while explicitly flagging low-confidence, non-standard clauses for human review rather than silently guessing.

Problem Statement

Product teams spend weeks running analyses that should take days. A business question gets asked Monday. The analyst spends Tuesday-Wednesday exploring data, Thursday writing SQL and building charts, Friday documenting findings, and Monday morning delivering a presentation. By then, the business context has shifted, the question has evolved, and the insight arrives too late to act on.

Solution

AI Analyst v2 is an AI-powered research partner that handles the execution layer while you keep the judgment layer. You ask a business question. The system runs an 18-agent pipeline that frames your question into testable hypotheses, analyzes your data with automated exploration, segmentation, cohort analysis, and root-cause investigation, validates findings across four layers (structural, logical, business rules, paradox detection), tells the story with a narrative arc and Storytelling with Data charts, and builds the deck: a branded, presentation-ready slide deck with speaker notes and export-ready assets.

Problem Statement

Marketing teams waste significant time creating, testing, and optimizing ad campaigns across multiple platforms (Google, Meta, LinkedIn, TikTok). Current workflows are fragmented, requiring manual ad copy creation, targeting configuration, budget allocation, and performance monitoring across disparate tools. This inefficiency leads to missed optimization opportunities, delayed market response, and reduced ROI focus. Teams need a unified, intelligent system to accelerate campaign creation and optimization.

Solution

Ad Genie is an AI-powered advertising automation platform that streamlines the entire campaign lifecycle. Using Claude AI and machine learning, it auto-generates compelling ad copy variants, intelligently configures targeting and budgets, monitors performance in real-time, and continuously optimizes campaigns for maximum ROI. The platform integrates seamlessly with Google Ads, Meta, LinkedIn, and TikTok, creating a unified advertising control center that reduces manual work and accelerates campaign velocity.

Problem Statement

Modern teams struggle with scattered project management tools that lack intelligent collaboration features. Teams lose alignment across multiple platforms, lack AI-powered insights into priorities and dependencies, spend time context-switching between tools, and can't quickly query their backlog or create issues from natural language. Project tracking becomes an overhead burden instead of a tool that accelerates shipping.

Solution

ProjectPilot AI is an AI-powered project management platform combining visual Kanban boards, real-time collaboration, and an intelligent AI assistant. Features include drag-and-drop Kanban boards with unlimited columns, AI assistant that understands your entire backlog and creates issues from natural language, real-time sync powered by Supabase Realtime, secure multi-tenant team workspaces with granular role management (owner, admin, member), and simple flat-rate pricing per workspace (not per seat). The AI knows every issue, priority, and assignee, enabling instant answers to "What are the urgent issues?" and "Who has the most work assigned?"

Problem Statement

Researchers face overwhelming information overload, with critical knowledge scattered across papers, clinical trials, conferences, and competitor data that is slow and hard to synthesize. The challenge was to let researchers quickly query all these sources in one place and get sourced, real-time answers in accessible formats.

Solution

Built ScienceAI, a Gen AI platform that lets researchers chat with scientific literature (papers, clinical trials, conferences, and competitor data) in one place. A Langflow RAG pipeline with a GraphDB and GPT-5-mini streaming (on Next.js 16) delivers sourced, real-time answers, while ElevenLabs TTS converts insights into on-demand audio summaries and dual-voice podcasts for faster, more accessible research consumption.

Problem Statement

Businesses receive high volumes of customer reviews that are slow and costly to read, prioritize, and respond to manually. The challenge was to automate review classification, prioritization, and response so teams could manage feedback at scale while improving customer satisfaction.

Solution

Built an LLM-powered solution that automates the classification and processing of customer reviews, predicting categories, assigning priority, suggesting actions, and generating sentiment-based responses. The solution streamlines review management, enhancing customer satisfaction and operational efficiency in handling feedback and inquiries.

Problem Statement

Businesses face growing security and productivity risks from spam SMS messages that can carry phishing and cyber-attack attempts. The challenge was to automatically classify incoming SMS texts as 'spam' or 'ham' to protect users and prevent attacks at scale.

Solution

Built a text-classification model that analyzes SMS messages labeled 'spam' and 'ham' to extract meaningful signals and predict whether a new message is spam. The solution automates spam detection, helping prevent phishing and cyber attacks before messages reach users.

Problem Statement

Hotels and booking platforms lose significant revenue to last-minute cancellations with little ability to anticipate them in advance. The challenge was to identify the key drivers of cancellations and predict at-risk bookings early enough to shape profitable cancellation and refund policies.

Solution

Built a predictive model that identifies the factors most influencing booking cancellations and flags which bookings are likely to be cancelled in advance. The solution enables data-driven cancellation and refund policies, helping reduce revenue loss and improve operational planning.

Problem Statement

In a competitive landscape, companies struggle to continuously gather, analyze, and act on customer feedback fast enough to drive innovation and growth. The challenge was to streamline feedback analysis so teams could surface actionable insights quickly and at scale.

Solution

Built an AI-powered customer feedback analysis solution that automatically ingests, categorizes, and analyzes feedback to surface themes, sentiment, and priority issues. The platform turns unstructured feedback into clear, actionable insights, enabling teams to continuously gather, analyze, and act on customer signals.