Current Dforgia client work, plus systems our founding team built in senior technical roles before Dforgia. Client and company names are withheld; the work and results are real.
Pilot with proof
Dforgia client workAgentsMultimodal
Agentic geocoding for hard-to-locate addresses
Pilot engagement · Logistics-tech startup
Problem
Free-text addresses full of landmarks, housing societies and gate-versus-building ambiguity defeat standard geocoding APIs, so deliveries and field visits go to the wrong place.
Approach
An AI agent that works the way a human analyst does: it generates cleaner address variants, cross-checks several geocoding sources, and verifies candidates against street-level and satellite imagery before choosing a location.
Result
A side-by-side comparison against standard geocoding, address by address, with clear reasoning for every location chosen and a per-address cost estimate for running at scale.
Live in production
AgentsGenAI
Multi-agent AI coaching platform
EdTech · Product development
Problem
A learning-science coaching product had to move from research concept to a secure, user-facing system that product, UX and subject-matter experts could all build on.
Approach
Turned the research into production requirements and architected a multi-agent coaching system with two production services: a coaching data API and a standalone service that lets non-engineers add new coaching modules.
Result
Deployed to production, with metric logging designed in from day one for analytics and model evaluation.
Manual work automated
AgentsGenAI
Email, document and helpdesk agents
Business automation
Problem
High-volume business email and internal knowledge questions were handled manually, slowing teams down.
Approach
Built a multi-agent email automation system, a retrieval pipeline over PDFs, spreadsheets and text files (question answering, summarisation, translation), a text-to-SQL data agent and an internal helpdesk agent.
Result
Manual document review and reply drafting replaced by agents, with people approving where it matters.
+24% over baseline
MultimodalGenAI
Multimodal document quality evaluation
Content platform
Problem
Document quality had to be judged at scale, including layout and visual content that text-only models miss.
Approach
Designed a fusion model combining vision-language embeddings with text signals, and built the evaluation protocol from scratch.
Result
Integrated into the enterprise workflow.
+40% over baseline
MultimodalPredictive ML
Multimodal forecasting model
Platform analytics
Problem
Forecasts built on tabular data alone were missing signal held in images and text.
Approach
Combined image, text and tabular features through embedding fusion, then tuned the architecture for accuracy.
Result
A substantially more accurate forecast than the existing approach.
75% ROUGE-1/L
GenAI
Fine-tuned small LLM for keyword generation
Education content
Problem
Keywords for large volumes of educational content needed to be generated cheaply and consistently.
Approach
Fine-tuned a small open model with supervised fine-tuning instead of relying on a large paid API.
Result
Strong agreement with human evaluators at a fraction of large-model cost.
83% across 11 categories
GenAI
Efficient toxic content classifier
Trust & safety
Problem
User content had to be screened across many harm categories quickly and cheaply.
Approach
Trained a transformer classifier with a frozen decoder for fast inference, and tested label granularity and loss functions.
Result
Accurate multi-category screening suitable for production speeds.
Real-time alerts
Predictive ML
Real-time anomaly detection
Operations monitoring
Problem
Problems in live metrics were being noticed too late.
Approach
Time-series anomaly models served through an inference API, with automatic alerts sent to the team's chat.
Result
Issues flagged as they happen instead of after the fact.