Stakeholders & Trade-off Navigation
- Translate one technical situation into exec language and eng language without distorting either
- Work trade-offs on the accuracy/latency/cost/privacy/usability pentagon with explicit, defensible reasoning
- Say no to a stakeholder request while offering two real options
- Write status updates that build trust — especially when the news is bad
- Decide when and how to escalate without burning relationships
| Spaced-rep warm-up: due cards from Days 169–170 | 10 min |
| ELI5 + tech read; study the trade-off radar | 20 min |
| Guided: two-audience drill + three pentagon cases | 40 min |
| Practice: the Friday demand reply | 20 min |
| Project: stakeholder map & comms plan | 20 min |
| Quiz + flashcards | 10 min |
Builds on: Day 160 — AI system design & the quadrilemma · Day 162 — The FDE role · Day 165 — Proposals & architecture docs
A good interpreter at a peace summit does not translate word-for-word — that produces nonsense and occasionally war. They carry the MEANING across: what this side fears, what that side needs, what both can live with. And they never editorialize; both sides trust them because the meaning arrives intact.
An FDE interprets between two languages every day. The executive speaks outcomes: money, risk, dates, headcount. The engineer speaks mechanisms: latency budgets, eval scores, retrieval precision. "The reranker adds 400 ms but lifts groundedness eight points" means nothing to a VP — but "answers take a beat longer and are wrong noticeably less often; wrong answers are what your compliance team flagged" is the same fact in their language. The second half of the job is that most requests are secretly trade-offs. "Make it faster AND more accurate AND cheaper" is like asking the summit for "everything, concede nothing." Your job is to put the real menu on the table — here is what each choice costs, here are two packages I recommend — so the people who own the outcome can actually choose.
Interview loops for FDE and applied-AI roles almost always include a stakeholder scenario: "the customer demands X, engineering says no, what do you do?" — because the day-to-day job is exactly this. Projects rarely die of bad code; they die of surprised stakeholders. The engineer who sends the Tuesday update that names the risk before it lands, and who answers "can it be faster?" with a priced menu instead of a promise, is the one customers ask for by name — and the one whose renewals close.
The five-way tug-of-war — two honest designs, one decision
step 1 / 5| design A: big model + rerank | design B: small model + cache | |
|---|---|---|
| sketch | frontier LLM, cross-encoder | distilled model, semantic cache |
Every AI system pulls on five ropes at once: accuracy, latency, cost, privacy, usability. You cannot max all five — engineering is choosing WHERE to lose. Two candidate designs for the dispatcher assistant:
Guided practice
One truth, two audiences
18 minThe situation, straight from your capstone's world: last Tuesday your LLM provider upgraded the default model version. Your regression gate (Day 141) caught it: groundedness on the golden set dropped from 92% to 84%, and two customer users have already reported "weird answers." You've pinned the previous model version as a stopgap; a re-tuned prompt for the new version is 3–4 days out. The pinned version is deprecated in 60 days, so staying put is not an option.
- Write the ENG update first (8–12 sentences): mechanism, eval evidence with numbers, the stopgap, the plan, what reviews you need.
- Compress to the EXEC update (max 5 sentences): what users saw, business risk, what you did within the hour, decision/date, what you need from them (nothing? say so).
- Self-check the exec version: zero unexplained jargon (would a CFO stumble?), at least one number tied to a consequence, a date, and the risk stated before they could hear it elsewhere.
- Read both aloud. If the exec version takes over 45 seconds, cut again.
Three trade-offs on the pentagon
22 minWork each case with the four-step method: name the tension → quantify both sides → package 2–3 options → make a recommendation but leave the choice. Write each as a half-page memo. Use your real capstone numbers where you have them (Days 140, 155, 156); estimate honestly where you don't.
Case 1 — accuracy vs latency: The champion user says answers feel slow. Your reranker (Day 117) costs ~500 ms of the 2.8 s p50 and buys +9 points of context precision. Options to consider: drop it, rerank only low-confidence retrievals, cache more aggressively, stream the answer so perceived latency falls (Day 155).
Case 2 — privacy vs quality: The CISO wants zero data sent to any external model. Local open-weights model scores 11 points lower on your golden set (or estimate from Day 103/128 experience). Options: hosted with a DPA + redaction (Day 132), region-pinned hosted endpoint, local model with the quality delta stated, hybrid routing by document sensitivity.
Case 3 — cost vs usability: Finance flags the monthly LLM bill. Cheapest lever is capping answer length and disabling streaming — which users will feel immediately. Better levers exist (Days 129, 156): routing easy queries to a small model, semantic caching, prompt-prefix caching. Package a cut that saves ≥30% with the least felt pain.
On your own
The Friday demand
20 minEmail from Meridian's VP of Claims, Thursday 4:52 pm: "Board meeting moved up — I need the assistant handling claims-status questions by Friday EOD. IT says the Snowflake connection is approved. Make it happen."
Reality: the Snowflake integration (Day 169's matrix) is 2–3 weeks of work — schema mapping, permission model, evals for a new answer type. Nothing safe ships by tomorrow.
Write the reply. Constraints: send-ready tone (this VP forwarded your last email to the CEO); no bare no; at least two concrete options with dates — think wizard-of-oz / hardcoded-scope moves from Day 166; one recommendation with a reason tied to THEIR outcome (the board meeting, not your backlog); under 200 words.
Hints: what could a board see Friday that is honest? A scripted demo on real-but-frozen data is not lying if you label it. What single sentence protects you from "but you demoed it, ship it"?
Stakeholder map & comms plan for the capstone
Write docs/stakeholder-plan.md as if the capstone is deployed at Meridian. (1) Stakeholder map: economic buyer, champion, daily users, blocker(s) — Day 163's roles — each with: what they care about, the metric they watch, their language (exec/eng), update cadence and channel. (2) The standing weekly update template (done/next/risks/asks) pre-filled with one realistic example. (3) An escalation path: three named triggers (e.g. "regression gate red > 48 h", "security finding severity ≥ high", "scope change touching the contract") and who decides what at each. (4) Your three trade-off memos from guided work, linked as appendices. This document is the communication half of Day 174's full simulation.
Common mistakes & misconceptions
- Translating jargon into vaguer jargon. "We improved retrieval precision" is still eng-speak; the exec version names what a human saw and what it costs or saves.
- Hiding the risk section until the risk lands. The Tuesday update that says "eval scores dipped, mitigation running, decision needed by Friday" builds more trust than a month of green — surprises, not problems, kill relationships.
- Answering a trade-off request with a promise. "We'll make it faster and keep accuracy" without measurements is a debt you'll pay with interest; put the priced menu on the table instead.
- Saying yes to the Friday demand and quietly cutting corners. The corner you cut (evals, permissions) is precisely what explodes in front of the board; no-with-options exists for this moment.
- Escalating the complaint instead of the decision. "The customer is being unreasonable" gets you sympathy; "here are the two options and who must choose by Wednesday" gets you a decision.
- Optimizing four vertices and forgetting usability. A private, cheap, accurate, fast assistant that requires three clicks and a VPN nobody has… gets zero queries, and zero times anything is zero.
Q1. The best EXEC rendering of "the cross-encoder reranker adds 500 ms p50 but raises context precision 9 points" is:
Q2. A stakeholder demands a feature by Friday that safely needs three weeks. The professional move is:
Q3. Which vertex did Day 160's quadrilemma gain today, and why does it dominate?
Go deeper — curated resources
- Crucial-conversation patterns for engineers — Practice the "facts → story → ask" pattern: lead with observable facts, label your interpretation as interpretation, end with a specific ask. It de-escalates almost any tense customer thread.
- Exec and eng updates written from the same incident; exec version ≤ 45 seconds aloud
- Three trade-off memos with quantified tensions and ≥ 2 options each
- Friday-demand reply under 200 words with two dated options
- Stakeholder plan committed; quiz ≥ 2/3
← Back: Day 163 named these stakeholders during discovery; Day 165 wrote for the mixed audience; Day 160 gave you four of the pentagon's five vertices. Today you practiced running the room they all sit in.
Forward →: Day 173 arms the money vertex with a real cost model — trade-off memos get sharper when the dollars are real. Day 174's objections are stakeholder pressure in written form, and Day 180's Demo Day presentation is the exec rendering of your whole 180 days.
Unlocks: D174 FDE Simulation II — Full Cycle