Deep dive · deployment friction is the bottleneck
The Forward Deployed Engineer toolkit
Core AI products are built in pristine environments; customer reality is messy. This console carries the eight modules an AI FDE — or the instructor teaching them — uses on the front line. Every module is deterministic and fully simulated: no API keys, no network, safe to run inside a locked-down client environment.
$ python -m fde_toolkit
FDE deep dive
The written briefing: the role, the four cognitive features, runtime guardrails, the skill stack and a day in the life.
$ read before the first session
SEOS deployment terminal
Cognitive mesh bootstrap, Neural ODE signal tuning, QFT/QSVM diagnostics and the deterministic G-LOCK failsafe test.
$ python -m fde_toolkit seos
Cohort management
Roster with skill levels, mentor-mentee pairing that prioritises stuck participants, and skill-mixed breakout rooms.
$ python -m fde_toolkit cohort
Environment diagnostics
Linux, Git, Docker, Kubernetes and network checks — each failure paired with the workaround that recovers the lab.
$ python -m fde_toolkit diagnose
Ontology mapping
Map TBL_CUST_99X_REV and friends onto a semantic knowledge graph, then query it in Cypher.
$ python -m fde_toolkit ontology
Memory architecture
Scratchpad, episodic and semantic tiers with eviction — plus a context rot curve that shows why it matters.
$ python -m fde_toolkit memory
Tool synthesis (MCP)
Turn an OpenAPI spec into scope-gated MCP descriptors and let the agent synthesise its own API calls.
$ python -m fde_toolkit tools
Intent routing
A fast classifier picks the specialist agent, with the cost, latency and hallucination trade-offs priced out.
$ python -m fde_toolkit router
Agentic ReAct loop
Thought → Action → Observation with MCP-style tool descriptors and a live context-window meter.
$ python -m fde_toolkit react
DAG orchestration
Decompose an AML investigation into agent tasks, topologically sorted with the critical path exposed.
$ python -m fde_toolkit dag
RAG lab
Chunk, embed, retrieve. Compare four chunking strategies and watch the lost-in-the-middle effect.
$ python -m fde_toolkit rag
Policy as code
Deny-overrides AML/KYC guardrails evaluated against a transaction, with the equivalent Rego shown.
$ python -m fde_toolkit policy
Audit trail
Every agent step and policy decision captured for regulatory replay, exportable as JSON.
$ python -m fde_toolkit audit
Retro log
Capture delivery friction as it happens and export the post-cohort report within 48 hours.
$ python -m fde_toolkit retro
Suggested day arc
taught session → lab → debrief
- 09:00Welcome + introductionsSurface the skill distribution, then open the Cohort module and build the pairs live.
- 09:30Environment checkWalk the diagnostics list in parallel with the room. Timebox failures to 5 minutes.
- 10:15Agent loopsTeach ReAct on the agent player. Step through one action at a time, narrate the context meter.
- 11:30Task decompositionMove to the DAG module. Show what runs in parallel and where the critical path sits.
- 13:30Enterprise contextRAG lab: change the chunking strategy mid-demo and let the retrieval ranking shift on screen.
- 15:00Runtime guardrailsPolicy module: run the structuring scenario, show four rules firing, reveal the Rego.
- 16:15Compliance replayOpen the audit trail and replay the session as a VP of Compliance would review it.
- 16:45DebriefLog every issue in the retro module before anyone leaves the room.
FDE vs other technical roles
| Role | Environment | Objective | Mindset |
|---|---|---|---|
| Software Engineer | Internal / HQ | Build scalable core product features | Optimal and maintainable for all users |
| Solutions Architect | Pre-sales / design | Design architecture that wins the deal | Prove it can work in theory |
| Data Scientist / ML | Internal / labs | Train models, reduce loss | Improve the metric |
| Forward Deployed Engineer | Client front lines | Make it work in this client's mess | Show ROI by Friday |
“I am not here to build a perfect system; I am here to build a system that works perfectly for you right now.”
Skill stack covered
the full-stack integrator