RAG and custom LLM applications
Answers grounded in your own documents, with the source cited on every response.
Postgres + pgvector, Pinecone or Weaviate · Claude, GPT-4 family or open-source models · eval harness before release
AI & AUTOMATION
AI in the workflow, not in a demo.
Retrieval-augmented LLM applications, intelligent document processing, computer vision and predictive models, built into the systems that run your business and reviewed by a person wherever the decision needs one. Under ISO/IEC 27001 and ISO/IEC 42001 controls.
How it works
How a document becomes a decision: six stages, two of which can stop it.
The same six stages hold for a RAG application or a vision model: retrieval or inference in place of extraction, the same validation, the same review gate, the same audit trail. That is what makes an AI system auditable under ISO/IEC 42001, and what a reviewer sees when they ask how a decision was made.
- Intake (Ingestion). Documents arrive from email, a portal or a scanner; each is fingerprinted and logged before anything reads it. Tools: S3, Blob Storage or Cloud Storage, Queue per source.
- Extract (LLM and vision). Fields are extracted with a confidence score per field, not per document, so a single doubtful value does not hide inside a confident page. Tools: Claude or GPT-4 family, Layout-aware vision models.
- Validate (Rules). Every extracted value is cross-checked against the source system: totals, identifiers, dates, the record it claims to belong to. Tools: Business rules as code, Lookups against the system of record. Gate: A failed check goes to review, never through.
- Review (Human-in-the-loop). A reviewer sees the field, the region of the page it came from and the confidence score, and decides. Tools: Review queue, Decision recorded against the reviewer. Gate: A low-confidence field needs a human decision.
- Write back (Integration). The approved record lands in the ERP, LIS, CRM or core system through its own API, idempotently, so a retry never duplicates it. Tools: REST or GraphQL, Idempotency keys.
- Audit (Observability). Every extraction, decision and write is traceable to a person or a model version, with the evals re-run on production traffic. Tools: Tracing, Regression evals, ISO/IEC 42001 controls.
What we build
Six things we build, and what each one integrates with.
Intelligent document processing
Invoices, claims, lab reports and contracts read into structured fields, with a confidence score per field and a review queue.
LLM and vision extraction · validation against the source system · write-back through its API
Computer vision
Medical imaging, manufacturing QC and document vision models trained on your data, deployed at the edge or in your cloud.
PyTorch · SageMaker, Vertex AI or Azure ML · drift monitoring
Predictive ML
Fraud, credit-risk and demand-forecasting models with explanations a reviewer can read.
Feature pipelines · SHAP explanations · MLflow model registry
Workflow automation
The steps around the model: routing, approvals, exceptions and the audit log.
Temporal or Step Functions · human-in-the-loop gates · exception queues
AI operations
Monitoring, evals, retraining and model governance after go-live.
Regression tests on production traffic · monthly governance review · ISO/IEC 42001 controls
How we work with you
Start with one workflow. Everything after it is optional.
- 014–6 weeks
Targeted automation pilot
One workflow, one system, automated end to end on your data and measured against a threshold agreed on day one.
You leave with
A working pilot, a measured accuracy figure and a go/no-go in writing.
Talk to an AI engineer - 028–16 weeks
Production AI build
A RAG, vision or ML system taken to production: integrations, review gates, monitoring and the audit log.
You leave with
A system in production that your compliance team can trace end to end.
- 03Ongoing
AI operations retainer
Monitoring, retraining, evals and a governance review every month.
You leave with
An engineer on the system, and a monthly record of how it behaved.
Five phases, with the artefact a production AI system needs at each one.
The full process- 01
Discovery & Strategy
1-2 weeks
Workflow map, a data-access inventory and the accuracy threshold you will hold us to
Signed off by you, before engineering starts
- 02
Architecture & Design
2-3 weeks
Solution architecture with the review gates drawn in, checked against ISO 27001 and DPDP controls
Signed off by your security team
- 03
Agile Development
4-12 weeks
The system, its eval set and the code review trail
Signed off by a senior engineer, every PR
- 04
Quality Assurance
2-4 weeks
Eval report against the agreed thresholds, with the failure cases listed
Signed off by your QA and compliance teams
- 05
Launch & Evolution
Ongoing
Model and drift monitoring, with a weekly digest
Signed off by your team, every week
A build engagement runs 9-21 weeks to launch, then ongoing. Every phase ends with a named artefact and a named sign-off, under ISO/IEC 27001 and ISO 9001 controls.
Recent work on this line
Two engagements running on this line now.
Healthcare / Diagnostics
Active engagementAI-driven reporting workflow for an Indian diagnostics centre
- Diagnostics
- Report generation
- NABL workflow
Life Sciences
Active engagementBioinformatics and AI-driven workflow engineering for an Indian biotech firm
- Biotech
- AI workflows
- Platform engineering
How we work
Built like a product company, shipped like one.
- 01
Audit-engineered by default
Every engagement runs under ISO/IEC 27001, ISO 9001, ISO/IEC 20000, and ISO/IEC 42001 controls. DPDP-aligned, with HIPAA / GDPR / RBI overlays available per project.
- 02
Founder-led delivery
Your discovery call is with a founder. The architecture review is with a senior engineer who stays on the project. No body-shop, no offshore handoff, no account-manager translation layer.
- 03
Bias to ship, not slide
We don't write demos that can't survive a Friday production deploy. Every milestone produces an artifact your team can use immediately: code, diagrams, telemetry.
Questions buyers ask
Straight answers, before the first call.
Is this retrieval-augmented generation, fine-tuning, or something else?
Usually retrieval-augmented generation over your own documents, because it keeps answers grounded and citable and it changes the day your documents change. We fine-tune when a task needs a style or a format the base model cannot reach with retrieval alone, and we say which we recommend and why in the discovery brief.
How do you measure whether it works?
An eval set is agreed in discovery, with the accuracy threshold written down before anything is built. The pilot is measured against it, and the same set runs as a regression test on production traffic after go-live, so a model update that lowers accuracy fails a test rather than a customer.
Can our documents stay in India?
Yes. We deploy into your own AWS, Azure or Google Cloud account in an India region, handle personal data under the DPDP Act, and can run open-source models inside that account when documents cannot be sent to a model provider.
What happens when the model is unsure?
Every extracted field carries a confidence score. Fields below the threshold go to a review queue where a person sees the field, the source region and the score, and decides. Nothing is written back to your system of record until the check has passed or a person has approved it.
Do you work with the systems we already have?
That is the point of the line. The model is the smaller part; the integration with your ERP, LIS, CRM or core system through its API, and the audit log around it, is where the work is. Nothing is replaced to make room for AI.
Insights
Reading on production AI.
Engineering notes on production AI, written by the engineers on the work.
Have an AI automation project in mind?
Book a 30-minute discovery call. We'll review the workflow and tell you honestly whether AI is the right tool for it.
+91 912-195-7728Hyderabad, IndiaEvery brief gets a senior review. Reply within 1 business hour, 9 AM-7 PM IST.