Agentic systems for domains where being wrong is expensive.
Drug regulation · medical education · venture diligence · skills policy
Language split measured across all 60 repositories, public and private.
Almost everything I ship is a variation on one problem: take messy institutional input — clinical dossiers, syllabi, pitch decks, inspector notes, meeting audio — and turn it into structured, auditable output that a human can defend in a review.
flowchart LR
A["Messy input<br/>dossiers · syllabi · decks · audio"] --> B["Extract<br/>and normalise"]
B --> C["Evidence layer<br/>every claim linked to a source"]
C --> D["Agent loop<br/>bounded tool use · web verification"]
D --> E["Scored, auditable output<br/>defensible in review"]
D -.->|"contradicted or low confidence"| C
E --> F["Audit log<br/>every AI decision, replayable"]
The interesting engineering is never the model call. It is the evidence graph underneath it, the decay function on stale signals, and the refusal path for when the agent is wrong.
| Project | What it is | Stack | Scale |
|---|---|---|---|
| VenIQ private | Agentic venture intelligence. Fact-checks startup claims through autonomous web research and scores them across 8 readiness frameworks (TRL/BRL/IRL/SRL…) into one living Venture IQ. Every metric graded Verified / Partial / Contradicted against extracted evidence, with scores that decay as signals go stale. Multi-tenant RBAC, audit logging, permissioned investor portal. | Next.js · FastAPI · Postgres · Claude + Gemini | 467 commits |
| Pre-incubation super-app private | One platform merging six separate products — assessment, ecosystem intel, mentor and manager consoles, innovator showcase, student onboarding — behind a shared sprint engine and role model. | Next.js · Prisma · Supabase | 216 commits |
| Boring Games private | A gamified operating system for a venture studio: quests, leaderboards, company tree, fundraising and segment finance, operative access tiers, briefing and timeline surfaces. | Next.js · Drizzle · Fly.io | 184 commits |
| Venture Analyst → Entrepreneurial OS private | Founder-facing ideation co-pilot with generative UI — a bounded Anthropic tool-use loop that surfaces interactive cards: matched resources, launchable tools, proposed sprints, and deep-research nodes that return a verdict on an assumption. | Next.js · Prisma · MCP · Anthropic | 107 commits |
| Project | What it is | Stack | Scale |
|---|---|---|---|
| SyllabusMapper private | Curriculum intelligence. Maps course outcomes to an L0–L5 industry skill ladder, computes gap analysis against job roles, and reverse-engineers a skill taxonomy from a job description. Ships its own MCP server and model registry; runs fully offline against a seeded local database. | FastAPI · Next.js · Gemini · MCP | 132 commits |
| KSDC District Skill Dashboard private | Skill-development dashboard for Dakshina Kannada district — institution and centre-of-excellence auditing with duplicate detection across a messy government dataset. In production. | React 19 · Vite · Supabase | 77 commits |
| ExamHermes private | Cross-platform exam-prep app built against college-specific syllabus specifications. | Flutter · Dart · Fly.io | 61 commits |
| erp_agent | Institutional ERP expanded for AICTE/NAAC compliance — 15 mandatory reports, each bound to an accountable role (TPO, librarian, anti-ragging cell, SC/ST cell…). | FastAPI · React | 37 commits |
| NitiSetu | Regulatory intelligence for CDSCO, India's drug regulator. Deterministic pseudonymization for DPDP Act 2023, cosine-similarity de-duplication of serious adverse event reports, gap analysis against CT-04 templates, full audit logging. | Next.js · FastAPI · hybrid LLM + local NLP | 12 commits |
| Project | What it is | Stack | Scale |
|---|---|---|---|
| mnema | Context engine for AI work — publishes a team's docs, decisions and workflows to any MCP-aware client, so agents answer from your context rather than their training data. | TypeScript · MCP | 13 commits authored |
| Forage private | Script-to-asset pre-production agent. Breaks a video script into visual beats, sources footage from legally usable archives only, license-classifies every item, and emits a manifest, CREDITS.md, credits.srt and an EDL. The download gate lives in the worker, not the UI — it never asserts fair use on your behalf. |
Python · Docker | 57 commits |
| AutoExtract private | Decision-driven extraction engine — searches, crawls, scrapes and normalises web data into structured datasets automatically. | FastAPI · Next.js · Docker | 43 commits |
| Dockmaster private | A policy engine in front of SSH, exposed over MCP. Claude Code chat, a CLI and a web panel are all clients of one rulebook, so none of them can diverge from it. Destructive operations are not discouraged, they are refused — denylist, read-only verb classifier, repo jail, single-use TTL-bound consent, secret masking. | TypeScript · MCP · ssh2 | 24 commits |
| Kiln | MIT reference lab for Generative UI patterns — a template per strategy (static, declarative, open-ended, MCP-UI), a BYOK bench to run them side by side, and "the Jury", a harness scoring generated UIs on accessibility, token adherence, structural validity, sandbox safety and latency. | TypeScript · pnpm · multi-provider | 4 commits |
| Project | What it is | Stack | Scale |
|---|---|---|---|
| Dreamer Cult private | Gamified social platform for project collaboration — share blueprints, peer-verify work in "The Arena", build reputation on a gamified profile. | Flutter · Supabase | 49 commits |
| ADHD companion private | A capture → decide → focus loop with planner, journal and insight surfaces. Offline-first sync, error tracking, shipped to production. | Next.js · Drizzle · PowerSync · Sentry | 41 commits |
|
If the system produces a score, a claim or a summary, it links back to the source that produced it — and grades how well that source actually supports it. |
If an operation is dangerous, the server refuses it. A warning in a README is documentation, not a control. |
Offline-first paths and local fallbacks, so a suspended hosted dependency does not take the whole product down with it. |
Languages
Web and mobile
Backend and data
AI and agents
Infrastructure
2,288 contributions in the last year — 2,093 of them in private repositories.
The public graph shows a fraction of the work.
The public graph shows a fraction of the work.