GTM intelligence systems
Company, contact, technographic, public-web, and client-proprietary data turned into structured account intelligence. Currently past 440,000 enriched cells written into client CRMs.
Applied AI Engineering · for production teams
I build production AI systems end-to-end.
Chris GarreApplied AI · Systems
Remote · United States
Remote · United States
I lead AI/ML at No Fluff Selling, where I design, build, and operate production AI systems for GTM teams. Alongside that work, I take on a small number of non-conflicting fractional builds for teams with expensive research, enrichment, routing, or decision workflows. The work is full-stack, schema-locked, and built to survive handoff.
Selective capacity
fractional builds only
fractional builds only
Default signal
I build the part of the AI system that has to survive contact with production.
The strongest fractional fit is usually a single workflow with measurable drag: company research, meeting prep, triage, routing, or structured decision support. I scope it tightly, build it end-to-end, and keep the operational surface explicit.
Likely work
- Source-grounded outputs with citations or audit trails.
- Strict schemas, observable failure modes, and human-readable exceptions.
- Integrations into the tools where the work already happens.
Relevant proof
Send the workflow you want made reliable.→- FastAPI services around model workflows, not model demos.
- Authenticated browser automation where APIs stop short.
- Post-handoff operation, alerts, and maintenance cadence.
Applied AI capabilities
AI research agents
Multi-stage research workflows for seller profiling, prospect research, account analysis, contact selection, and GTM recommendations.
Schema-locked outputs
Structured reports that downstream apps, CRMs, and workflow automations can render and trust without hand cleanup.
Workflow orchestration
n8n, Python/FastAPI, async callbacks, queues, retries, throttling, progress polling, and cost controls. The largest orchestrator I run is 92 nodes and has cleared 5,056 executions at 99.6% success.
Full-stack product surfaces
Next.js interfaces for generation, admin review, usage monitoring, API cost visibility, and pipeline operations.
Vector retrieval
Qdrant collections, OpenAI embeddings, per-prospect stores, deterministic preprocessing, and evidence-grounded synthesis.
Cost control
Deterministic gates in front of every model call, so the LLM never runs on a record that was going to be filtered anyway. That holds third-party data and model spend to a few cents per record.
Selected case studies
012026
Multi-agent GTM research
Multi-agent account research system
A production multi-agent system that turns a seller URL and prospect URL into an evidence-grounded account-research report in 3-5 minutes. Five n8n workflows coordinate seller profiling, prospect research, adaptive retrieval, account analysis, contact selection, and batch callbacks.
n8nFastAPINext.jsPostgresQdrantGeminiOpenAI
Manual work4-8hrreplaced
Run time3-5min
StatusIn production
022025
Wealth-tech sales intelligence
Sales-intelligence pipeline
A production prospect-enrichment pipeline that turns a name and email into a 22-field structured advisor profile, factual bio, and sales-persona tag synthesized from six independent sources across the RIA / broker-dealer hierarchy.
PythonFastAPIPlaywrightPydanticOpenAIHubSpotFly.io
Output22fields
Sources6+feedsreconciled
StatusIn production
032026
Proptech incident signal
Incident signal pipeline
A single 92-node orchestrator turns third-party incident webhooks into qualified, contactable opportunities in a CRM. Ownership resolution for trusts and LLCs, enrichment in cost order, a versioned severity model, and structured record creation in four to eight seconds.
n8nHubSpotPostgresOpenAIfly.ioNext.js
Executions5,056runs99.6% ok
Data + model2¢/run
Conversion5x lift
042026
Panamanian legal corpus
Citation accuracy over a scanned archive
Attribution over decades of scanned government legal documents, where the identifier is the first thing scanning destroys. Type-aware line-anchored matching cut misattribution from 36.00% to 0.69%. My own measuring instrument then flattered itself twenty-fold before I caught it.
PythonPostgresOCRMeasurement harness
Number-only locate36.00% wrongtype-aware 0.69%
True misattribution6.40%instrument said 0.34%
StatusPre-launch
Operating background
No Fluff Selling
Production GTM intelligence systems, multi-agent research workflows, retrieval, strict schemas, admin surfaces, and operational controls.
Immortal Oceans AG
In-silico modeling, biotech data work, and preclinical/clinical strategy for a drug development program.
OceanBuilders
Built and operated a 3D-printing PPE workshop and secured a $196k grant during the COVID supply crunch.
PanaSea + field operations
Aquaculture strategy, investor materials, grant work, and restoration operations where evidence, constraints, and execution all mattered.
Recent writing
2026 · 07Measuring what a model refuses to answer4 min2026 · 04The strict-schema discipline, and what it costs you2 min2026 · 03Scrape, integrate, or ask: a decision sketch4 min2026 · 02Six-source synthesis - notes from production4 min2026 · 01On being the only engineer on the project4 min2025 · 11Playwright in production: the boring infrastructure4 min2025 · 10What I owe a system after handoff4 min
Get in touch.
For focused AI systems architecture or production workflow work, write directly. I reply within two business days. Best fit is a bounded fractional build where the business workflow is expensive enough to justify senior engineering.
Directhello@chrisgarre.com
Cadence2 business days
EngagementsFractional · 6-12 wk
CapacitySelective · 1-2 d/wk
Set in IBM Plex Sans Condensed · IBM Plex Mono[ CG · 001 ] · © 2026 Chris Garre · Build 2026.07.27