Call Happens
A customer talks to an agent through Amazon Connect.
AWS account 465572510980 · us-east-1 · deployed MVP
From Call to Answer
A clear pipeline for turning customer service calls into evidence-backed operational insight: redacted transcripts land in S3, a 15-minute compiler analyzes them with Bedrock and memory distillation, Athena stores the facts, and the Codex skill answers questions with charts.
Redacted Transcript Lands
Contact Lens writes redacted `VOICE` JSON to S3. The stack does not read raw audio.
15-Minute Compiler Runs
EventBridge starts Step Functions. New objects are discovered using a watermark.
Bedrock Applies Meaning
Labels from the taxonomy: driver, root cause, intent, friction, escalation, confidence, evidence.
Memory Is Distilled
Calls are rolled up into hourly, daily, and weekly summaries: issue types, changes, claims, and evidence.
Athena Stores Facts
Conversations, utterances, annotations, evidence, and memory versions land in Iceberg tables.
Codex Answers
The skill queries Athena, reads memory pages, finds evidence, and generates dashboards.
M
What “memory distillation” means here
Bedrock analyzes each call one at a time. Memory distillation is the next layer up: the compiler takes those call-level labels, evidence spans, durations, and counts, then rolls them into time-window summaries. The goal is to maintain a useful memory for each hour, day, and week: what people called about, what changed, which examples prove it, and what an LLM should know before answering. Athena keeps the structured facts; the distilled memory gives Codex, agent assist, or a future customer portal compact context without rereading every transcript.
Real Demo Example
One Delta synthetic call, end to end.
This example uses `syn-delta-airline-v1-0242-0cb34882`, a synthetic redacted Contact Lens call from the Delta demo dataset. It shows what arrives, what labels get applied, what gets distilled, and what Athena can answer.
1. Source Transcript Snippet
S3 redacted JSON
Customer
I am calling about mobile check-in loop. I already tried the mobile app and I still need this fixed today.
Agent
Thanks for calling Delta. I can help with that. I will verify the account and look at what happened.
Customer
The main problem is voucher not applying. This is affecting my trip timing.
Agent
I see the digital channel failure case. Let me check the recent activity and policy notes before I make a change.
2. Labels Applied
convintel.annotation.v1Bedrock’s role is not just “summarize.” It turns the conversation into structured facts: taxonomy labels, issue intent, root cause, friction and escalation signals, confidence, and evidence spans that point back to redacted turns.
3. Distilled Assessment Page
hour/day/week memoryDistilled memory: Daily issue summary for Delta demo
Input: individual call annotations for the time window. Each call contributes labels, evidence turns, duration, sentiment, resolution status, and a source contact ID.
Reduction: group calls by taxonomy label, compare this window with earlier windows, preserve representative examples, and turn the facts into a narrative memory that says what an LLM should know about the current operating picture.
Output: a versioned memory page for an hour, day, or week. A future LLM can use it as context: "Delta callers are currently over-indexing on digital channel failures, especially vouchers not applying; here are the counts, trend direction, representative contacts, and evidence."
- Conversations annotated: 250
- Top driver: digital_channel_failure: 79
- Top root cause: voucher_not_applying: 34
- Next: identity_verification_failure: 25
- Memory version: dd1a6058421ef2540ef5e898637979bd
4. Final Athena Facts
query-ready tables| Table | Example Row |
|---|---|
ci_conversation | contact_id `syn-delta...0242`, synthetic `true`, duration `516,882 ms` |
ci_utterance | turn `turn-0003`, role `CUSTOMER`, redacted content about voucher not applying |
ci_llm_annotation | driver `digital_channel_failure`, root cause `voucher_not_applying`, confidence `0.90` |
ci_memory_page_version | driver summary page for `delta-airline-v1` |
What drove calls this week?
Which root causes are accelerating in the last three days?
Show representative contacts for voucher failures.
Generate a dashboard of driver mix and evidence.
Drill Down
How the simple process maps to AWS components.
Click a component to see what it does, why it exists, how it scales, and where it lives in AWS.
What It Solves
From transcript piles to evidence-backed operating intelligence.
The compiler gives operations teams a repeatable way to ask what is driving calls, why issues are moving, and which contacts prove the pattern, without opening raw recordings.
Driver Discovery
Normalizes every conversation into comparable drivers, root causes, friction points, and resolution signals.
Evidence Discipline
Annotations carry utterance-level evidence spans, memory pages, and representative contact IDs for defensible summaries.
Demo Reset
Synthetic datasets can be generated, reseeded, validated, ingested, and reset by `dataset_id` without touching real data.
Analyst Speed
Codex can query metrics, find contacts, read memory pages, and generate dashboard artifacts from the packaged skill.
Redaction Boundary
The data plane is scoped to redacted Contact Lens JSON. Original audio and unredacted transcripts are out of bounds.
Future Channels
The schema leaves space for chat transcripts, BI dashboards, agent-assist surfaces, and customer-specific taxonomies.
250Delta demo conversations
7,027Redacted utterances
15 minIncremental schedule
11Iceberg tables
Runtime Flow
The same seven-step story, with the technical gears exposed.
The non-technical view is call to answer. The technical view is Step Functions coordinating Lambda, Bedrock, Glue, S3, and Athena.
1. Call and redaction
Amazon Connect and Contact Lens produce redacted voice analysis JSON. Raw audio is outside this workflow.
2. Discovery every 15 minutes
EventBridge starts the compiler, which checks S3 for new redacted files using an incremental watermark.
3. Normalize for analytics
The Lambda turns Contact Lens JSON into stable conversation, utterance, and signal rows.
4. Bedrock applies taxonomy
Bedrock labels driver, root cause, intent, resolution, friction, escalation, channel conflict, novelty, confidence, and evidence spans.
5. Distill memory
The compiler compresses call-level labels into a reusable assessment page that summarizes patterns across the run.
6. Load into Athena
Glue merges staged facts into Iceberg tables so analysts can query trends and examples.
7. Answer through Codex
The skill asks Athena, reads memory pages, finds evidence contacts, and renders charts or dashboards.
Technical drill-down: state machine
The deployed state machine is `tsf-convintel-dev-compiler`. It runs
`DiscoverNewObjects`, exits through `NoNewObjects` when object count is zero, and otherwise
proceeds through `Normalize`, Glue loads, `Annotate`, `CompileMemory`, and optional KB sync.
Glue steps retry transient concurrency errors.
Technical drill-down: idempotency
Iceberg writes use table-specific merge keys such as `conversation_id`,
`conversation_id + utterance_id`, and `version_id`. Replayed runs update existing rows instead
of duplicating contacts.
Technical drill-down: redaction guard
Real data must arrive under `/Redacted/` and pass Contact Lens shape validation. Synthetic
data is explicitly marked with `AnalyticsMetadata.Synthetic` and `synthetic=true`.
Technical drill-down: memory distillation
This is the Karpathy-like compression step: the system takes many call-level facts and
turns them into hour/day/week memory pages. Those pages become context for Codex,
dashboards, agent assist, and future systems so they can answer with the latest
operating picture without rereading every transcript.
Scale and Performance
Designed for incremental work first, batch backfills second.
The current deployment favors safe, observable increments. The next performance step is materializing common dashboard aggregates into `ci_trend_snapshot`.
Bounded Discovery
`MAX_DISCOVERY_KEYS` limits a Lambda discovery pass. Watermarks prevent repeatedly scanning old objects.
Iceberg Merge
Glue 5.0 writes Parquet-backed Iceberg tables and merges by deterministic keys for replay safety.
Athena Result Reuse
The helper enables recent-result reuse for repeated dashboard and metric queries when Athena can reuse a matching result.
Concurrency Buffer
The Glue loader allows up to five concurrent runs, and Step Functions retries transient Glue concurrency errors.
Annotation Control
Bedrock calls are isolated to the annotation phase, making cost and latency easier to measure and tune.
Materialized Trends
`ci_trend_snapshot` and memory pages are ready for precomputed hour/day/week windows, so dashboards and LLMs can reuse the same distilled operating picture.
Integration Model
One compiler, many consumers.
Codex is the first analyst surface, but the data contract is intentionally generic: Athena tables, S3 evidence, and memory-page versions can feed BI, agent-assist, or customer portals.
Consumer
Codex SkillAnswers, charts, dashboards.
BI DashboardAthena metrics and snapshots.
Agent AssistKnown issues and policy memory.
Customer DemoResettable synthetic datasets.
Interface
`convintel.py`Query helpers and HTML/SVG output.
Athena SQLGoverned table access.
Memory PagesVersioned insight summaries.
Generator CLIGenerate, validate, ingest, reset.
Evidence
Contact IDsRepresentative records for summaries.
Trend RowsDriver/root-cause windows.
Evidence SpansUtterance-level rationale.
Dataset MetadataSeed, prompt, model, quality scores.
Codex Skill Today
Uses Athena, S3 memory pages, offline Delta snapshot fallback, and generated dashboards.
BI Tomorrow
Point QuickSight or another BI layer at `ci_trend_snapshot` and aggregate views.
Knowledge Base Later
Enable Bedrock KB once an OpenSearch Serverless vector collection and index are selected.
Data Model
Conversation facts, annotation facts, memory lineage.
The model keeps transcript turns separate from annotations and evidence so analysts can aggregate safely while still drilling down to representative contacts.
| Table | Grain | Primary Use | Join Path |
|---|---|---|---|
ci_conversation | Contact/conversation | Header, dataset, source, queue, duration. | Root table by `conversation_id`. |
ci_utterance | Transcript turn | Redacted content, role, offsets, sentiment. | Many to `ci_conversation`. |
ci_llm_annotation | Conversation annotation | Driver, root cause, confidence, novelty. | One to `ci_conversation`. |
ci_annotation_evidence | Evidence span | Quote and rationale supporting an annotation. | Annotation to utterance. |
ci_memory_page_version | Compiled page version | Versioned insights and dashboard provenance. | Dataset and page lineage. |
ci_trend_snapshot | Trend window | Materialized dashboard acceleration. | Dataset, driver, root cause. |
ci_claim | Analyst claim | Claim text, SQL, confidence, evidence contacts. | Links claims to contacts and pages. |
Technical drill-down: annotation schema
`convintel.annotation.v1` includes drivers, intent, root cause, resolution status, friction
points, escalation triggers, channel conflict, novelty score, confidence, and evidence spans.
Technical drill-down: synthetic data contract
Synthetic datasets emit Contact Lens-shaped redacted JSON under
`conversation-intelligence/synthetic/source/<dataset_id>/Analysis/Voice/Redacted/`.
Ground-truth labels are evaluation-only and not treated as model findings in normal demos.
AWS Links
Jump directly to the deployed resources.
Links open the AWS Console in `us-east-1`. Access depends on the active AWS account and IAM session.