What we build
Production AI systems across seven verticals — designed, built, and deployed by our team.
01 —
Real-time AI over live video streams. Deployed in warehouse and logistics environments.
AI that watches your cameras so your team doesn't have to.
A production AI surveillance platform that runs over existing CCTV infrastructure. Detects PPE compliance violations, theft behaviors (hand-in-pocket detection), and safety anomalies in real time — without replacing cameras. Deployed across multi-site logistics operations with 20+ cameras per site.
PPE detection — helmets, vests, harnesses, footwear — real-time alerts on violations
Behavioral anomaly detection. Hand-in-pocket theft pattern recognition with temporal tracking.
Works over RTMP streams on existing cameras. GPU-accelerated via DeepStream + TensorRT.
Search surveillance footage the way you search Google. Find any moment, instantly.
A video intelligence system that indexes surveillance footage and lets operators search for specific events, objects, people, or behaviors using natural language queries. Ask "show me all moments someone entered through the rear door after 10pm" — and get timestamped clips, not raw footage to scrub through.
Natural language query over hours of indexed footage. Results in seconds.
Frame-level semantic indexing. Objects, behaviors, identifiers — all queryable.
Timestamped clip extraction. Shareable, audit-ready evidence packaging.
02 —
Agentic AI workflows that bring transparency and automation to clinical and administrative processes.
Search. Reason. Deliver. AI workflows that work through complex documents and produce structured, transparent outputs.
A configurable agentic workflow framework built around a single repeatable pattern: ingest a trigger document, search across a corpus of relevant documents, reason through the evidence, and produce a structured output — with full transparency into every step taken. The same engine powers different use cases by swapping the domain, the document corpus, and the output format.
One live example: insurance claim denial appeals for hospitals — the agent ingests a denial notice, retrieves clinical notes and policy documents, reasons through the grounds for denial, and produces a submission-ready appeal letter with cited evidence. Same pattern, different domain: prior authorization support, contract review, clinical documentation gap analysis.
Agentic RAG pipeline · Multi-step LLM reasoning · Document retrieval · FastAPICapabilities
Retrieves the right documents from any corpus — clinical notes, policies, contracts, guidelines
AI reasons step-by-step through retrieved evidence. Every inference is traceable and explainable.
Structured final output — letter, report, summary, decision — with source citations throughout.
03 —
Enterprise AI analytics for drug discovery, clinical data, and research operations — deployed on Databricks.
AI-powered analytics over clinical and research data — built for enterprise pharma on Databricks.
A production AI analytics layer deployed for a top-10 global pharmaceutical company on Databricks. Enables research and ops teams to query complex clinical and research datasets in natural language, surface insights from structured and unstructured data, and accelerate decision-making across drug discovery and trial operations.
Databricks · Delta Lake · Python · LLM pipeline · Enterprise governanceCapabilities
Built natively on Databricks. Integrates with existing data lakehouse and Delta Lake.
Natural language querying over clinical datasets. Structured and unstructured data fusion.
Enterprise-grade — built for the data volumes, compliance, and governance requirements of global pharma.
04 —
AI for content transformation, real-time stream monitoring, and automated learning material generation.
Any document. Any format. Turned into a short-form learning video — automatically.
An agentic pipeline that ingests any file — PDF, Word, PowerPoint, video, audio — and converts it into a structured short-form learning reel. Designed for corporate L&D teams, compliance training, onboarding, and edtech platforms that need to scale content production without scaling headcount.
PDF, DOCX, PPTX, MP4, MP3 — any format in, structured learning reel out
AI identifies key concepts, sequences them pedagogically, and scripts the reel
Short-form video with narration, visual structure, and learning checkpoints
AI that watches live streams and surfaces what matters — in real time.
A real-time AI monitoring system for live video streams — built with an OBS plugin integration. Analyzes video and audio in real time to detect content policy violations, surface viewer engagement signals, flag anomalies, and provide streamers and platform operators with live intelligence. Applicable to brand safety, trust and safety, and streamer performance analytics.
Real-time analysis of live video and audio. No post-processing lag.
OBS plugin — works with existing streamer setup. No infrastructure changes.
Content flagging, engagement signals, anomaly detection — all surfaced live.
05 —
Vision AI that reads architectural documents and converts them into structured, renderable data.
Feed in a 2D architectural PDF. Get structured, 3D-renderable JSON out.
A VLM-powered pipeline that ingests 2D architectural floor plan PDFs and extracts structured data — rooms, walls, dimensions, units, common areas — into a JSON schema consumable by 3D rendering engines. Eliminates weeks of manual digitization for architecture firms, construction managers, and real estate platforms.
Qwen VLM · OpenCV · FastAPI · EC2 · Three.js output schemaCapabilities
Walls, dimensions, rooms, units, common areas — all extracted with linewidth-aware analysis
Output is structured JSON — ready for Three.js, BIM, or any 3D rendering pipeline
Multi-page, multi-unit floor plans. Processing accuracy targeting 85–90% post-POC.
06 —
End-to-end AI inspection platform — turning drone footage into structured, actionable asset intelligence.
Upload drone footage. Get a structured inspection report. No manual review required.
Pyrack designed and built an end-to-end drone inspection platform — from the web application to the AI pipeline running underneath it. Drone footage is ingested, processed through a defect detection and anomaly classification pipeline, and converted into structured inspection reports. Eliminates manual footage scrubbing and removes the need for physical access to hazardous or hard-to-reach infrastructure.
Computer vision · Defect detection AI · FastAPI · Full-stack web platformCapabilities
Drone video and image ingestion across sites, assets, and inspection runs — structured intake pipeline
AI identifies defects, structural anomalies, and surface degradation — with location and severity tagging
Structured inspection reports — annotated images, findings summary, severity ranking, audit-ready output
07 —
Conversational intelligence, data automation, and monitoring tools for alternative investments, private credit, and financial reporting.
Ask your database a question in plain English. Get an answer in seconds — with a chart.
A multi-agent agentic AI layer that translates natural language into precise SQL queries, runs them against your databases, and returns cited, chart-ready answers — with no technical knowledge required. Supports multi-database setups, PDF document Q&A, hybrid answers, and conversation follow-ups.
LangGraph · FastAPI · PostgreSQL · AWS Bedrock · ChromaDB · PgBouncerCapabilities
Capital calls, distributions, statements, fund/investor info — all queryable in plain English
Upload PDFs — notices, agreements, statements — and ask questions from them
Three-layer sensitive data protection. Role-based access control. Full query audit trail.
Eliminate the 1.5-week monthly reporting cycle. Generate complex investment performance reports in hours.
Reports in hours, not weeks. Incremental compute — only recalculates what changed.
Formal adjustment layer. Every change logged, attributed, timestamped, reversible.
Config-driven report templates. Format changes without code deployments.
Upload financial documents. Compute metrics. Flag exceptions automatically.
Trial Balance, P&L, Income Statement, Balance Sheet, Cash Flow — structured extraction
Configurable formula engine. Leverage, DSCR, ICR, LTV, NAV ratios — all automated.
Continuous threshold monitoring. Breach detection and exception surfacing for ops teams.
AI-generated intelligence profiles on companies and decision-makers — built for B2B sales teams.
Funding stage, headcount signals, tech stack, recent news, growth indicators
Role history, mandate signals, recent activity, likely pain points — per decision-maker
Pushes structured profiles into your CRM. Works with Gmail OAuth2 and outbound sequences.