tokencostguard / docs /case-study-template.md
lixiaowww
CSP: allow huggingface.co frame-ancestors
249c849
|
Raw
History Blame Contribute Delete
2.61 kB

TCG Case Study Template

Use after a beta user agrees to be referenced (anonymized or named). Keeps ROI claims defensible for marketing and investors.

File naming: artifacts/case-studies/YYYY-MM-<slug>.md


Metadata

Field Value
Customer (public name)
Segment indie / YC startup / SMB product team
Stack LangChain / AutoGen / custom / other
Agent topology # agents, avg steps per task
TCG plan Free / Pro
Period analyzed YYYY-MM-DD → YYYY-MM-DD
Author

1. Baseline (before TCG)

  • Monthly LLM spend (USD):
  • Primary models:
  • Main workload: e.g. coding agent, support bot, RAG pipeline
  • Pain: e.g. bill spike, no attribution, unknown cache usage
  • Tools used before TCG: LangSmith / none / spreadsheet / other

2. Integration

  • Endpoint: POST /api/v1/optimize (+ feedback if used)
  • Frequency: per trace / nightly batch / other
  • Data sent: full logs / truncated / PII redacted by customer
  • Applied recommendations? yes / partial / no (critical for honesty)
Intervention Applied? How
Compression / HCA summary
Routing tier change
Prompt structure / cache blocks (v1)
HERMES / policy

3. Metrics (quantitative)

Metric Before After Δ Source
Est. monthly tokens (input) TCG / provider bill
Est. monthly cost (USD)
Avg tokens per successful task
% calls on premium model
Cached input tokens (if provider reports)
p95 latency (optional)

TCG-reported fields (from API):

  • _guard.baseline / _guard.saved (note: may be estimated from logs, not provider invoice)
  • _vmm.saving_ratio if compression applied
  • _routing.targetTier if used

Claimed savings %: ___%
Method: (baseline_cost - after_cost) / baseline from TCG or invoice comparison — state which.


4. Qualitative

  • What surprised the team?
  • What they would not give up (quality guardrails)?
  • Would they pay $12/mo? Why / why not?

5. Limitations (required)

  • Savings include model routing, not only prompt optimization
  • Sample size: N traces / M days
  • Seasonality or product change during period
  • TCG did not proxy LLM traffic (beta)

6. Publishable quote (optional)

“…” — Role, Company


7. Internal only

  • tenant_id / trace ids (do not publish)
  • Support issues / blockers for product