AI X DEV. GEN AI
Agentic systems that read documents, retrieve knowledge, and improve themselves
Our GenAI Agentic solutions leverage multi-agent workflows, RAG architectures, and intelligent document extraction systems to process complex real-world data. From scanned PDFs and mixed-format documents to ambiguous user queries, our agents continuously learn from feedback and user corrections to deliver smarter, more accurate outcomes.
20+
agentic systems in production
3
document formats supported
12+
local + API models combined
100s
correction passes per extraction
What We Build
From Research To Production Outcomes
Multi-agentic prompt optimization
Our GenAI Agentic solution powers a self-improving extraction workflow where every user correction is automatically fed back into the system. Intelligent agents identify recurring patterns, optimize extraction logic, and refine prompts continuously, enabling higher accuracy and efficiency without the need for manual tuning.
- Every user correction is stored as a paired example of what the model extracted versus the approved value
- A supervisor agent reviews correction history to find recurring patterns and fields causing friction
- A refinement agent updates extraction instructions to handle the specific formats and edge cases seen in production
- New prompts are validated against historical examples before deployment to avoid regressions
LangGraph
LangSmith
MongoDB
BlokSeT mediator
Classifies modification requests as Valid, Forbidden, or Needs Approval by retrieving and reranking rules from a PDF rulebook, with conversation memory and structured escalation reports for engineering.
- Blends embedding similarity with LLM-based reranking, weighted by domain such as safety, electrical, mechanical, or cosmetic
- Vague requests are forced into Needs Approval and trigger clarification questions before classification
- Embedded conversation history fast-tracks near-identical repeat requests using prior classifications
- Escalations include similar past cases, engineer guidance, and a documentation checklist, all linked to Sort case IDs
LLaMA 3.1 8B via Ollama
ChromaDB vector stores
Two stage reranking
Three outcome routing
TTF co-pilot
Searches across PDFs, spreadsheets, presentations, and Word docs in one query, preserving the exact page, slide, or sheet location for every chunk and bringing scanned legacy documents into the same searchable corpus.
- Every chunk is tagged with its page, slide, or sheet location, surfaced directly in retrieval results
- PDFs are sampled to detect scanned pages, routing only true scans to the VLM and keeping standard extraction fast
- PowerPoint files use a manual slide-by-slide extraction path to keep slide numbers accurate
- Unified search across mixed engineering datasets
- Material type (copper or aluminum) is inferred from file paths, letting engineers filter results by conductor
Kreuzberg extraction
Qwen3-VL for scans
ChromaDB and Qdrant
location-aware chunks
CML coordinate extraction
Derives exact 3D (E, N, EL) coordinates from scanned piping isometrics, replacing manual transcription with a multi-pass vision pipeline handed off to a deterministic geometry engine for numerical accuracy.
- Reads every printed coordinate block on a drawing, then validates topology against EL-checksum labels with up to six correction passes
- CML bubble detection combines classical CV geometry with a separate VLM tag-reading pass, taking the union of both
- Corrects for image downsampling drift and fits a 2x3 iso-projection matrix from trusted anchor and EL-check nodes
- Handles the full range of AG-series drawing types in one pipeline, with no per-sheet configuration needed
Claude Vision API
EL checksum validation
Affine projection
Velosi-compatible JSON
Electrical knowledge base: agentic RAG
A multi-agent system that reasons about its own knowledge scope at query time, routing LV distribution and LV protection questions to specialist agents that decompose, retrieve, grade, and retry rather than relying on a single fixed collection.
- A coordinator agent reads knowledge base descriptions at runtime and routes each question, with no hardcoded mapping
- Specialist agents decompose compound questions, retrieve per sub-query, deduplicate, and grade the result before retrying
- The compliance agent only cites clause numbers and values that appear verbatim in retrieved context, with iterative hallucination guards
- Conversation history persists across server restarts, and a three-layer test suite measures hit rate, MRR, and coverage
Agno multi-agent team
local qwen3 via Ollama
ChromaDB MMR retrieval
SQLite session memory
Roof quantity learning agent
Learns a roof takeoff procedure once through a guided conversation with an expert, then applies the same procedure automatically to any new architectural roof plan.
- A five-question guided session with a subject-matter expert is compiled into a reusable standard operating procedure
- Vision model reads scale notation, building dimensions, and counts roof components by symbol shape, returning strict JSON
- Separates vision uncertainty from arithmetic by extracting drawing data first, then applying stored rules deterministically
- New drawing types are supported by running a new teaching session without touching existing code
GPT-4o vision
Acontext skill memory
Two-stage extract and calculate
FastAPI endpoints
Core Tech Stack
Built With Proven AI & Engineering Technologies
Real-World-Impact
From Research To Production Outcomes
Manual → Agentic
Rule lookups and document triage that took a person minutes per case now run as automatic, routed decisions.
1 Hours → API call
Coordinate transcription from piping isometrics replaced by automated extraction with built-in validation.
One session, many runs
A single guided teaching conversation with an expert produces a skill reused indefinitely on new drawings.
No documents excluded
Scanned legacy files are routed to a VLM and join the same searchable corpus as digital-native files.
Verdicts with citations
Compliance checks return structured outcomes with clause-level references, traceable back to source rules.
6 auto-corrections
Checksum loops run up to six correction passes before flagging an extraction as approximate.
How It Works
From Raw Data To Production AI
Every engagement follows the same path: audit, engineer, deploy, and continuously improve.
Data Audit
Map every source, assess sensor coverage and determine treatment per signal.
Label Correction
Algorithmic boundary fixing so the model learns from actual asset behavior.
Feature Design
Physics-informed features, rolling windows and validation checks.
Model Build
Training, evaluation, threshold tuning and explainability baked in.
Deploy & Serve
FastAPI, Docker and Kubernetes with batch or real-time inference.
Feedback Loop
User corrections automatically feed retraining pipelines.