Production AI systems, multi-agent frameworks, edge vision models, and full-stack software, architected from concept to deployment.
18 systems shipped
Agentic AI/Live · v2.0
OnRamp
Step Pipeline
9
Languages Parsed
20+
Role Tiers
9
A repository is ingested once into a dependency graph plus an entity graph of classes, functions and API routes, and that index is reused everywhere. The autopilot classifies issues by difficulty, assigns each one to whoever currently holds the fewest active tasks, opens a PR per issue, re-parses the PR head to graph-diff it against the base, and emits a structured senior review. Because a merged PR advances the linked task and closes the originating GitHub issue, the loop closes without anyone touching a board.
An AI-native workspace that unifies notes, tasks, projects, and a knowledge graph. Notes, tasks, and projects are nodes with real edges, so the assistant can traverse relationships instead of treating every record as an island.
The whole sales lifecycle on a shared schema where every query is scoped by workspaceId: lead → contact → deal → site visit → unit → cost sheet and payment plan → booking → milestone collection → documents → reporting. The AI layer sits on top of that rather than beside it, with next-best-action, call analysis, revenue and collections forecasting, and a workspace-wide /ask grounded in the tenant's own rows and documents.
Ordered by where a tool sits in a real system rather than by popularity: serving, routing, retrieval, fine-tuning, agents, orchestration, guardrails, prompts, evals. The facts about a tool (deployment, licence, cost) live apart from its link, and the build throws if either exists without the other, so a wrong licence can be corrected without touching prose.
A multi-provider orchestration pipeline (triage, then a deep analyzer, then a cross-check validator with chain-of-thought negotiation strategy) turns a job offer into an actionable brief in seconds instead of hours.
A production fairness audit across intersectional demographic slices, with Pareto frontier plots and automatic drift detection. Built for EU AI Act, NIST AI RMF, and ISO 25059 compliance review ahead of pre-market clinical evaluation.
Five adaptive difficulty levels with a multi-provider fallback chain across Gemini, GPT, and Claude, so a provider outage degrades quality rather than taking the product down.
Real-time ingestion into Firestore with time-series forecasting and heatmap rendering. A hybrid statistical + ML approach beats pure ML once the data gets noisy.
A complexity classifier routes each request to the cheapest model that can handle it, with checkpointing and human-in-the-loop review. Precise routing plus bounded agent steps keeps the token bill predictable instead of open-ended.
Pluggable analyzers emit a unified schema so findings correlate across tools instead of landing in separate reports, with an LLM pass that explains the result in plain language.
TensorRT ONNX engine with INT8 quantization-aware training, GStreamer hardware-accelerated decode, and hand-written CUDA kernels. Histogram equalization handles the low-light and motion-blur conditions on real track footage.
Parses Mermaid, PlantUML, images, and PDFs into a graph, then fans out to seven scoring agents. The fallback chain runs Groq → NVIDIA → OpenRouter → Gemini → Ollama → Hugging Face, backed by an 18-rule heuristic engine that works with zero API keys.
A planner breaks the request into searches, an extractor normalizes noisy pages, and a grounded writer only claims what retrieval actually supports, every statement carries its source.
Feature engineering over transaction velocity and behavioral entropy, paired with a threshold-moving strategy so the classifier stays useful at a fraud rate of one in ten thousand.
GPT-2/4 regex pre-tokenization, byte-level fallbacks, and a vocabulary retrained on technical text, trading pure-Python speed for a measurable drop in token count on domain corpora.
A complete recommendation pipeline, feature extraction, model training, serving, and a product surface, rather than a notebook that stops at evaluation.