Contracts, not ratings

What I can be held to.

A progress bar labelled “LangChain 85%” tells you nothing. So every skill below is written as a contract, what goes in, what comes out, and paired with something shipped that you can go and open.

Agent orchestration

inA multi-step task and a set of available tools

↓

outA completed action, with the tool calls and reasoning visible

LangGraph/LangChain graphs and direct Claude tool-calling loops: deciding when a model should call a tool, what to do when it calls the wrong one, and how to keep the whole thing observable instead of a black box.

evidence

The assistant on this site, tool calls exposed →The MCP server it calls →

Agent-readable interfaces (MCP)

inA product or dataset that only humans can currently use

↓

outA typed MCP tool surface, machine-readable endpoints, and an llms.txt an agent can discover on its own

Exposing capability as tools rather than pages: schema design and length caps, read/write separation, registry publication, and plain-HTTP fallbacks for clients that don't speak JSON-RPC.

evidence

This site's MCP server, 4 tools →GET /api/resume.json →GET /llms.txt →

RAG pipeline design

inAn unstructured document set and a natural-language question

↓

outA grounded, cited answer, plus a number for how often retrieval actually hit

Chunking strategy, embeddings and pgvector, retrieval tuning, and measuring quality before and after rather than shipping on vibes. Built into Analytra to query 11 live data sources in natural language.

evidence

Analytra, AI analytics SaaS, built solo ↗How the assistant on this site grounds its answers →

LLM features users actually understand

inA model output your users can't act on, a rejection, a score, an error

↓

outA plain-language explanation and a guided path to the next step

Turning model output into something a non-technical user can act on, which is usually where AI features fail. Shipped in production inside a GDPR/CCPA-compliant flow at 100M+ user scale, where a confusing rejection means a lost customer.

evidence

This site's assistant explaining tool results in plain language →Shipped at 100M+ user scale (role detail) →

Workflow & orchestration engines

inA process hard-coded per customer, changeable only by shipping code

↓

outConfigurable multi-step workflows with dynamic rules, changed without a deploy

Architecting the engine rather than the one-off: configurable multi-step flows, dynamic rule evaluation, and automated paths that cut manual intervention. Built one that shortened enterprise customer integrations from weeks to configuration.

evidence

Role detail in the résumé →Analytra, multi-source pipeline orchestration ↗

Authentication & browser cryptography

inA product that needs strong auth without holding biometric data

↓

outWebAuthn/FIDO2 and OIDC flows, with biometrics processed on-device under homomorphic encryption

Shipped the Ultrapass Web SDK, WebAssembly biometric auth with fully homomorphic encryption, and implemented OIDC/OAuth 2.0 (PKCE + Authorization Code Grant). Co-maintained SimpleWebAuthn.

evidence

SimpleWebAuthn (open source, co-maintained) ↗

Leading delivery on a team

inA team of 8 and a platform enterprise customers depend on

↓

outArchitecture decisions made, engineers mentored, cross-functional delivery shipped

Lead Full Stack Engineer at Private Identity, owning architecture and delivery for a platform serving 100M+ users and supporting Google, CVS Health and Uber.

evidence

Essay: coding interviews test the wrong thing (8,800+ readers) →Experience timeline →

$ haroon.verify --all

Prefer to check for yourself?

Connect the MCP server to your own agent and interrogate the résumé directly, or read the raw JSON. No forms, no gate.

Connect an agent GET /api/resume.json