One WhatsApp voice note turns paddy straw into income, not smoke: how we built clearsky

def verify_signature(raw_body: bytes, header: str | None, app_secret: str) -> bool:
    if not header or not header.startswith("sha256="):
        return False
    expected = hmac.new(app_secret.encode(), raw_body, hashlib.sha256).hexdigest()
    return hmac.compare_digest(header.removeprefix("sha256="), expected)

self.t.put_item(
    Item={"wa_message_id": wa_message_id, "status": "processing",
          "ttl": int(time.time()) + TTL_SECONDS},
    ConditionExpression="attribute_not_exists(wa_message_id)",
)

InboundQueue:
  Type: AWS::SQS::Queue
  Properties:
    VisibilityTimeout: 720  # 6 × the processor timeout
    SqsManagedSseEnabled: true
    RedrivePolicy:
      deadLetterTargetArn: !GetAtt InboundDLQ.Arn
      maxReceiveCount: 3

s3.put_object(Bucket=s.media_bucket, Key=key, Body=data, ContentType="audio/ogg")
tr = boto3.client("transcribe", region_name=s.aws_region)
name = _job_name(msg_id)
out_key = f"transcripts/{name}.json"
tr.start_transcription_job(
    TranscriptionJobName=name,
    Media={"MediaFileUri": f"s3://{s.media_bucket}/{key}"},
    MediaFormat="ogg",
    LanguageCode=s.transcribe_language,  # "hi-IN"
    OutputBucketName=s.media_bucket,
    OutputKey=out_key,
)

body = s3.get_object(Bucket=s.media_bucket, Key=out_key)["Body"].read()
transcripts = json.loads(body)["results"]["transcripts"]
return " ".join(t["transcript"] for t in transcripts).strip()

- Effect: Allow
  Action: [transcribe:StartTranscriptionJob, transcribe:GetTranscriptionJob, polly:SynthesizeSpeech]
  Resource: "*"
- Effect: Allow
  Action: [s3:GetObject, s3:PutObject, s3:HeadObject]
  Resource: !Sub "${MediaBucket.Arn}/*"

idx = data.rfind(b"OggS")
granule = struct.unpack_from("<q", data, idx + 6)[0]
return round(granule / 48000, 2)

audio = polly.synthesize_speech(
    Text=speech[:2900],
    VoiceId=s.polly_voice,      # "Kajal"
    Engine=s.polly_engine,      # "neural"
    LanguageCode="hi-IN",
    OutputFormat="mp3",
)["AudioStream"].read()
key = f"media/out/{msg_id}.mp3"
s3.put_object(Bucket=s.media_bucket, Key=key, Body=audio, ContentType="audio/mpeg")
return presigned_get_url(s.media_bucket, key, 3600)

def build_tools(phone: str, ctx: TurnContext) -> list[Any]:
    """The agent's tools, bound to `phone`. The phone number is never an LLM-visible argument."""

    @tool
    def book_pickup(field_id: str) -> dict[str, Any]:
        return ctx.record("book_pickup", {"field_id": field_id}, impl.book_pickup(phone, field_id))
    ...

if p == "bedrock":
    return BedrockModel(model_id=s.bedrock_model_id, temperature=s.agent_temperature,
                        max_tokens=s.agent_max_tokens, region_name=s.aws_region)
if p == "openai":
    return OpenAIModel(client_args=client_args, model_id=s.llm_model_id,
                       params={"temperature": s.agent_temperature, "max_tokens": s.agent_max_tokens})

capacity_cond = "attribute_not_exists(booked_acres) OR booked_acres <= :cap_minus_a"

condition="#s = :offered AND (attribute_not_exists(expires_at) OR expires_at > :now)"

ReadAttributes: [email, custom:district, custom:buyer_id, custom:baler_id]
WriteAttributes: [email]

Auth:
  DefaultAuthorizer: CognitoJwt
  Authorizers:
    CognitoJwt:
      IdentitySource: $request.header.Authorization
      JwtConfiguration:
        issuer: !Sub "https://cognito-idp.${AWS::Region}.amazonaws.com/${UserPool}"
        audience: [!Ref UserPoolClient]

client.admin_update_user_attributes(
    UserPoolId=pool, Username=username,
    UserAttributes=[{"Name": ATTRIBUTE_FOR[app.role], "Value": entity_id}],
)
client.admin_add_user_to_group(UserPoolId=pool, Username=username, GroupName=GROUP_FOR[app.role])

Events:
  Daily:
    Type: ScheduleV2
    Properties:
      ScheduleExpression: cron(0 18 * * ? *)
      ScheduleExpressionTimezone: Asia/Kolkata