Autonomous agents
Harness and loop engineering that finishes the job
We build workflow-specific harnesses — not a generic chatbot — in front of an in-network model gateway. Agents gather, act, and verify, with a test gate before they call the work done.
- An in-network coding agent — single-file Python, about 1,100 lines, standard library only — with file operations, atomic multi-edit, diff preview, git operations, memory, skill packs, and session save and resume. A test gate sits in front of done.
- A 12-step LLM migration pipeline, about 12,000 lines, that rewrites a Java/Spring service — POMs, security configuration, controllers, permission evaluators, properties, Maven build, and git push — in about 2.5 minutes against about 7.5 hours by hand.
- Custom harness engineering and loop engineering on an enterprise LLM gateway: Azure versus Bedrock routing, reusable context, model-aware temperature, stop-sequence hardening, and evaluation harnesses. Gather → act → verify — never a silent template fallback. Token efficiency on the same workloads: 42% lower token spend, 2.4× useful completions per 1k tokens, and about 68% prompt-cache hit on repeated enterprise workflows.
42%
Lower token spend on the same workloads
2.4×
Useful completions per 1k tokens
~68%
Prompt-cache hit on repeated workflows
Agents and multi-agent systems
Incident intelligence, then a human on the last step
We put multi-agent systems on the telemetry, tickets, and runbooks you already run — and on the service call graph — then keep a person on the last step. Mean time to recovery drops 75%.
- Incident intelligence that reads telemetry, tickets, and runbooks together, then hands the on-call a ranked hypothesis and MCP-guided remediation with the evidence attached. 75% lower mean time to recovery.
- MCP-native agents on the tools your teams already use, plus a graph-intelligence platform: 15 MCP tools and 20 REST endpoints over a service call graph, with multi-turn session tracking, entity extraction, and intent routing to 12 handlers.
- A human keeps the last step. We do not run production operations fully unsupervised.
75%
Lower mean time to recovery
RAG and LLM architecture
Answers with a citation — and a refusal when there is none
Grounded retrieval over tickets, documents, and runbooks. Context graphs and Graph Architecture hold the relationships; evaluation harnesses measure whether the system stays honest.
- RAG and GraphRAG over the knowledge your teams actually wrote, with citations, a confidence signal, and a refusal when the source does not support the answer.
- A graph and LLM pipeline that writes wiki-ready API documentation and migration-eligibility assessments per service, including token-scheme classification checked against ground truth.
- Evaluation harnesses, guardrails, and regression suites treated as product surface. The same harness and gateway pattern — routing, reusable context, stop-sequence hardening — is how we hold 42% lower token spend and 2.4× useful completions per 1k tokens, with about 68% prompt-cache hit on repeated workflows. Context graphs and Graph Architecture sit under the retrieval. That is harness engineering for evaluation.
Data processing with AI
Schema, mapping, inventory, and warehouse rules
We replace the mapping grind and the warehouse-rule grind with pipelines that propose, score, and explain — and we inventory the estate so the work has a map.
- Schema and attribute mapping with confidence scoring and a review queue — up to 60% less time on that work, and 3× throughput on attribute mapping.
- AST inventory of a Java/Spring estate: 63 services, 236 controllers, 467 controller-level tasks, 506 REST endpoints. Lakes, ETL, Spark, and Databricks sit under the work.
- A manifest-driven rule engine — 118 externalized SQL detection rules against a 432-table gold warehouse — with SHA-256 hash-locked artifacts, fail-closed integrity, UUID5 idempotency, and 44,000+ records per run. GxP-grade, Part 11-style auditability as a capability.
60%
Up to 60% less time on schema and mapping work
3×
Throughput on attribute mapping
Platforms and delivery
Lakes, migration, regulated quality, production serving
The rest of the system: Graph Architecture for blast-radius and scoping, legacy-to-cloud extracts with lineage, serving-layer RBAC, and review-ready validation writing — on AWS, Azure, and Google Cloud.
- Static-analysis Graph Architecture: 75 Java microservices, 504 endpoints, 10 UI applications, and 15 external systems into Neo4j, with 88.9% edge resolution, used for blast-radius, migration scoping, and security review. That is graph engineering.
- Legacy-to-cloud and mainframe-class extracts with lineage preserved. Serving-layer RBAC — 5 capability modules, 11 role grants, a 25-endpoint registry. Sixteen immutable DDL releases. GAMP Category 5 validation writing as a capability — URS, system specification, architecture, and configuration — without naming a product. PlantUML held in lockstep with the design pack. Multi-cloud: AWS, Azure, GCP.
- Android and iOS product engineering. We designed and built the Ingre apps (ingre.ai): camera scan, label parse, and the harmful-ingredient call, for food and beauty labels.