Ahmad Rishi

Lead a 4–6 engineer AI team; own the agentic-platform roadmap and org-wide AI tooling across business units, and mentor engineers and new-grad hires to ship production features.

• Initiated and led the Agentic Development Board — an internal AI platform on AWS that drives a business request through spec → plan → code → tests → documentation as LangGraph multi-agent workflows with human approval gates between phases. Wrote the repo's first commit and led it as maintainer across ~8 contributors, merging 94 feature branches to master. ~116K LOC across a FastAPI backend and a Next.js 15 / React 19 / TypeScript frontend.
• Built the platform's LLM provider abstraction — one interface over two incompatible coding-agent runtimes (a forked CLI binary and an in-process SDK), swappable by a single Parameter Store value — including the runtime factory, the Claude Code SDK execution path, and the subprocess layer that makes forking a 600-second child safe.
• Designed 16 workflow types as a persisted phase machine over a 46-table MySQL schema, deliberately without a LangGraph checkpointer so one store answers "where is this task"; wrote the natural-language → security-validated read-only API workflow and its self-healing validate→repair loop, and the Kafka event publisher and SSE streaming path that surfaces multi-minute agent runs to the browser.
• Authored the multi-model agent framework — 16 model-specialized agent definitions (discovery, analysis, backend/frontend implementer, code-reviewer, impact-analyzer, test-writer and E2E variants) packaged as drop-in agent kits, 38K lines, sole author — with every convention derived by measuring the real repositories; plus the platform's Atlassian OAuth 2.0 + JWT session layer end to end, the auth foundation every integration on the platform authenticates against.
• Operated the platform's AWS deployment — VPC, autoscaling EC2 behind an ALB with ACM TLS, RDS MySQL, self-hosted Kafka (KRaft), ECR and SSM-managed secrets, provisioned with Terraform (44 resources across 26 AWS types) and released via containerised CI/CD.
• Initiated CodeHarbor and led its scan engine — an agentic LLM security scanner (FastAPI + LangGraph + Claude Code SDK) that runs a coding agent inside a sandboxed per-execution clone, scans GitLab diffs on a schedule against a 9-check security checklist and makes it prove every finding, with Pinecone similarity de-duplication and severity-based email + dashboard alerts — 207 commits and ~25K insertions of my own into the engine.
• Stood up the team's LLMOps backbone — a self-hosted LiteLLM gateway fronting every major provider behind one endpoint with per-key budgets and routing, and Langfuse (ClickHouse) tracing — then migrated production traffic off Claude 3.5 Sonnet to a candidate ~12× cheaper per token, gated on an LLM-as-judge replay harness over real conversations rather than on price alone.
• Standardized org-wide LLM tooling (Cursor, LangGraph, Langfuse) and built the adoption-analytics stack behind it: a custom VS Code/Cursor extension (diff engine) feeding a FastAPI/MySQL service and an Angular 20 + NgRx dashboard, plus a queue-backed prompt-logging service (Redis Streams, a 300 req/s rate ceiling, bounded queue shedding to 503, XCLAIM dead-worker recovery, dead-letter stream, idempotent batch upserts).
• Built and shipped a WhatsApp AI agent pilot on n8n — 22 workflows / 357 nodes with Twilio inbound orchestration, a 5-tool agent, and Redis conversation state (debounce list, durable attribute hash, 10-turn memory) including a stale-reply guard; load-tested it to destruction (30,300 requests ramped 26→940/s) and identified a hard ceiling of ~65 sustained req/s with 70.7% timeouts beyond it — successful throughput stayed flat regardless of arrival rate, proving a saturated worker pool rather than a network limit. Also authored the fleet-sync and backup tooling covering 9 n8n instances / 5,558 workflows.

• Built a production Jira/Confluence knowledge-management RAG (LangChain + LangGraph over Pinecone, dual knowledge bases; migrated from ChromaDB) indexing 30–60K Jira issues across an enumeration reaching 40 projects, answering with numbered source citations generated by construction so a citation cannot reference an unretrieved document. Blind pairwise evaluation on 283 real questions: preferred over the incumbent tool 240–30–13 by an LLM judge and 246–34–3 by a human domain expert, the two agreeing 91.5% of the time. Async Flask + Celery, CI-deployed, with Redis-cached rolling conversation summarisation, a DB-backed feedback loop and latency analytics.
• Shipped a second production RAG over the company's OpenAPI/Swagger specs using a custom OpenAPI-aware semantic chunker (endpoint/schema/tag chunks with $ref resolution) and a LangGraph ReAct retrieval agent with text-embedding-3-large.
• Designed a multi-provider, multi-judge LLM evaluation framework integrating OpenAI, Anthropic, Gemini and DeepSeek — 7 evaluation arms replaying real production conversations against the Claude incumbent, with deliberate reciprocal cross-judging and a self-preference probe — and used it to clear the migration to a candidate ~12× cheaper per token. Key finding: swapping the judge moved the pass rate 37.4 points on byte-identical outputs (82.7% → 45.3%) against a sampling error of ±9–11, and 37.6% of judge verdicts failed to parse.
• Built an automated code-security workflow (GitLab webhooks + LLMs) reviewing every commit diff across 15 repositories, emitting structured severity-ranked findings to email and an audit log for remediation.
• Delivered a passport CV/OCR pipeline — OpenCV (CLAHE contrast, denoise, perspective de-skew), EasyOCR, and PassportEye MRZ check-digit validation, benchmarked against DE-GAN/GFPGAN/NAFNet deep-enhancers — with an automated MRZ-confidence + image-clarity + alignment QA gate that cut the manual photo-review step for flagged passports.
• Built an event-driven OCR microservice (FastAPI + Kafka, Regula DocReader, Gemini image enhancement) with idempotent processing, dead-letter recovery and Prometheus metrics; plus a fastText language-detection API covering 200+ languages and an 85-language LLM quality-evaluation study.

• Built the Maids.cc Ziwo telephony integration suite: a production CTI screen-pop linking inbound/outbound calls to agent ERP profiles via a deterministic E.164 number-normalisation funnel, and a live missed-calls analytics dashboard over the Ziwo CDR API with adaptive time-window bisection around the 5,000-record export cap.
• Engineered a real-time waiting-agents queue counter (FastAPI, HTTP/2, uvloop) on the IVR's critical path feeding Ziwo smart-routing, and reverse-engineered Ziwo's storage API to snapshot 417 production call-flow/IVR configurations that had no export path.
• Owned and gated the ERP's Playwright + TypeScript E2E test framework as reviewer-integrator — merged every change into main across the library and its consumer template (16 Angular Material component wrappers handling CDK overlays and virtual scroll, 53 helper utilities, a parameterized-SQL database-verification client, 232 tests), drove it to a scoped release on a self-hosted private npm registry, and caught defects in review including a credential-leak gap in the ignore rules.
• Containerised and self-hosted two internal operations dashboards — moving one off managed serverless onto Docker Compose with a deployment runbook, hardening the image build so build-time secrets never reach an image layer, and acting as merge gate on the trust-boundary rewrite that moved every privileged database call server-side.

• Co-founded an early-stage startup and led all technology; built an AI-powered legal document-management platform (AWS + LangChain) that automated review and cut document-processing time significantly.
• Delivered a multilingual OCR + LLM extraction pipeline for legal documents.

• Built Python web scrapers (Scrapy) and browser-automation workflows (Playwright) for SMM platform clients, and delivered PHP/Laravel client websites with admin panels, REST APIs, and third-party integrations (payments, social, analytics).