A support agent that asks before it acts
customer-ops-agent: answers from a knowledge base and takes four real actions (ticket, CRM update, email, booking), every mutating one gated behind human approval.
Systems I've built
Thirty systems, from agents that ask before they act to a satellite-alert investigator. The full write-ups include the part that matters: the constraints, the decisions, and what failed first. Numbers are only quoted where they were actually measured.
A WhatsApp booking agent that cannot double-book
ConverseIQ: a multi-channel agent for a booking-heavy business that answers from the business's own records, books appointments, captures leads, and routes phone calls into the same brain.
Answers from your PDFs, with the source highlighted on the page
DocLens: a full-stack RAG chatbot over uploaded PDFs with hybrid retrieval, reranking, a grounding check that refuses to guess, a side-by-side viewer that highlights the cited page, and reference-free answer evaluation.
A financial research agent for the Pakistan Stock Exchange that shows its sources
FinTex: four agents over Qdrant documents, Supabase time-series and live PSX prices, fused into cited answers with a confidence score. The longest-running build here.
A price prediction API that explains every rupee
AutoPricer: a production-style ML service for used-car resale prices. SHAP attributions that provably sum to the prediction, JWT and API-key auth, a fail-open Redis cache, async batch scoring, drift monitoring, and a retraining gate that refuses ties.
Flagging invasive species a year before official detection
Beachhead: a multi-agent early-warning system that fuses citizen-science sightings, climate suitability, documented introduction pathways and propagule pressure into triaged field-survey recommendations, evaluated retrospectively on real GBIF data.
A RAG accuracy audit in a box
rageval: point it at any retrieval or agent system through a thin adapter and get retrieval recall, faithfulness and hallucination rate as a report, plus a regression harness with a CI gate so it never silently degrades again.
An LLM proxy your app adopts by changing one URL
llm-gateway: an OpenAI-compatible proxy with a two-tier cache, complexity-based model routing, prompt-injection screening, PII redaction, output validation and provider failover. The only change in your code is the base URL.
RAG that can answer multi-hop questions
NexusGraph-RAG: converts documents into a Neo4j property graph while keeping Qdrant vector embeddings, then fuses k-hop graph traversal with semantic search through reciprocal rank fusion for questions dense retrieval can't answer.
Adversarial testing for LLM apps, reported the way security teams read
AegisRedTeam: fires a corpus of adversarial payloads at an LLM endpoint or RAG pipeline, classifies each response with a two-stage evaluator, grades per OWASP LLM Top 10 category, and emits an executive PDF alongside JSON, HTML and Markdown.
Satellite change alerts, turned into evidence a newsroom can use
Sentinel Dossier: a five-agent pipeline that fuses near-real-time satellite change detection with legal land-use boundaries and multilingual news, reasons about whether a change is legally suspicious, and writes cited, human-reviewable dossiers for the cases that matter most.
A daily agenda built from your inbox, Slack and calendar
TaskPilot: four LangGraph agents read across Gmail, Slack and Google Calendar, extract genuine action items, merge cross-channel duplicates, score priority and effort, and block-schedule a realistic day around existing meetings.
Rollback for agents that touch the real world
agent-saga-engine: a transactional runtime that executes agent tool calls as a Saga, each forward step paired with a compensating inverse, so a failure at step 3 automatically undoes steps 2 and 1 instead of leaving them permanently committed.
Years of forest audio, turned into cited ecological assessments
Chorus Ledger: tracks a site's acoustic indices and species detections over time, separates real change from seasonality, detects change points, retrieves candidate explanations from the field log and literature, and writes a confidence-scored ledger entry that says what it couldn't determine.
Farming advice that knows when to hand you to a person
Rootstock: a farmer describes what they see; the agent structures it, pulls their soil profile and recent weather, retrieves vetted guidance filtered to their situation, writes an answer where every sentence cites a numbered claim, and then decides whether it should be answering at all.
A browser agent for the SaaS tools that have no API
VisionBot-RPA: an autonomous web agent that combines a vision LLM with accessibility-tree parsing to run multi-step workflows on sites without public APIs: form filling, data extraction, dynamic application interaction, from a natural-language goal.
Triage for dark fishing vessels, from radar return to cited brief
Bright Return: cross-references synthetic-aperture-radar vessel detections against marine protected area boundaries and vessel-risk data to produce prioritized, cited briefs for human review. The analyst layer on top of a mature detection problem.
A fine-tune that had to beat an honest baseline first
ForgeTune: fine-tunes a small open model to turn unstructured PSX company announcements into schema-valid JSON, on a hand-corrected dataset, evaluated against a prompted baseline instead of a straw man.
A voice assistant built like a small service
VoxAgent: wake word, offline speech recognition, a LangGraph tool-calling agent (weather, search, semantic notes, reminders) and spoken replies; the same core drives the mic loop, a text REPL and a live reasoning dashboard.
Small language models running entirely on-device
EdgeSLM-Engine: compiles Qwen2.5-0.5B and SmolLM2 to ONNX with KV-cache support, quantizes to INT8/FP16, and runs a hand-rolled generation loop on CPU, DirectML, CUDA, or in the browser via WebGPU. No cloud, no per-token cost.
Instruction datasets synthesized from raw documents
SynthData-Forge: chunks unlabelled text, generates grounded seed instructions, hardens them with Evol-Instruct mutations, produces gold responses with Claude, filters by reward-model score, dedupes with MinHash LSH, and exports JSONL/Parquet/HF with full lineage per sample.
Four recommenders, one ranked watchlist
ReelFusion: fuses transformer content similarity, matrix factorization, deep collaborative filtering and gradient boosting into a weighted ranking with per-signal score breakdowns, plus SVD fold-in personalization without retraining.
Clinical notes into validated structured data
ClinicalParse: an extraction pipeline that turns unstructured clinical notes into schema-enforced diagnoses, medications and dosages, with a self-correcting retry loop on validation failure, drug-interaction flagging, confidence annotations and FHIR-lite output.
Synthetic data for rare species, built so a null result can win
Ghost Trap: fine-tunes a diffusion model on a handful of rare-species camera-trap images, composites generated examples onto real backgrounds, and measures on a sealed real-only test set against a pre-registered decision rule whether recall actually improved.
Downscaling coral heat stress, validated against real loggers
Reef Grain: a conditional diffusion model generating ensembles of reef-scale thermal fields from 5 km satellite SST, validated against held-out in-situ temperature loggers. Headline finding: the satellite product is too smooth to contain the variability, and the project reports that number instead of hiding it.
TradingView alerts into risk-managed paper orders
OrderBlock: a webhook-driven execution engine that validates signed TradingView alerts against Smart Money Concepts rules, filters them through an ML signal-quality scorer, applies position sizing and daily limits, and places orders through an Alpaca paper-trading adapter.
One order, fanned out across every account, controlled from Telegram
OrderPilot: a Django/Celery execution engine that validates risk, sizes a trade per account, dispatches it idempotently across multiple brokerage accounts, and exposes real-time alerts and a remote kill-switch through a Telegram bot and web cockpit.
Clean 3D point clouds from handheld scanning rigs
PointFusion: fuses 200 Hz IMU data with 30 Hz depth and RGB streams to produce motion-compensated, spatially registered point clouds, exportable as PLY/PCD, served through a FastAPI job API.
Resume-to-job matching where every point is traceable
ResumeAlign: a calibrated 0-100 fit score built from a four-layer skill-matching cascade (exact, synonym, fuzzy, embedding) over Gemini-extracted structures, showing which requirement matched via which layer against what evidence, and which are genuinely missing.
An AI career OS that can't invent skills you don't have
CareerPilot AI: transparent job-match scoring, resume tailoring with an honesty guard, job aggregation from official ATS APIs, and an application tracker whose outcomes train a personal match model. Next.js, Postgres, credits and billing.
Nothing matches that combination. Try widening the filters.
Bring me the problem. Leave with how I'd build it.
A 20-minute architecture call. No pitch, no slide deck. If I'm not the right person for it, I'll say so and point you somewhere better.